AI Models

Meta has launched its most advanced open-weight artificial intelligence (AI) model to date, Muse Spark Open. This new model is designed to rival leading Chinese AI models such as Qwen and Kimi K3, which have dominated the market in recent years.

According to Meta’s chief of AI Alexandr Wang, Muse Spark Open is a significant development for the company. In a tweet, Wang described the model as ‘the most capable US open-weight model to date,’ with 816 billion total parameters and 42 billion active parameters. The model also boasts a 1 million-token context and native multimodal capabilities.

Muse Spark Open is an open-weight AI model, which means that developers can download its parameters or weights and customize them as needed. This approach differs from closed models like Anthropic’s Claude or OpenAI’s ChatGPT, where the underlying code is not accessible to users. Wang emphasized that Muse Spark Open ‘excels at long-horizon tasks, complex software engineering, and multi-step reasoning,’ making it a valuable tool for developers.

Meta has been working on improving its AI capabilities in recent years, with a focus on developing more advanced open-weight models. The company’s efforts have paid off, as Muse Spark Open is now considered one of the top contenders in the market. Chinese developers such as Alibaba and Moonshot AI have previously dominated this space, but Meta’s new model has closed the gap.

Muse Spark Open requires significant computing power to run efficiently, but users can create a ‘supercomputer’ on their own desktop by running the model on two Mac Studios with M5 Ultra chips. This collaboration between Meta and Apple is an unusual partnership that highlights the growing importance of AI in the tech industry.

Muse Spark Open, a new coding-optimized model developed by Meta, has been optimized to run efficiently on the NVIDIA GB300 NVL72 platform. This 816B mixture-of-experts (MoE) model boasts an impressive 42 billion active parameters and a massive 1 million token context window, allowing it to hold entire projects with complex dependencies in memory without breaking tasks into smaller pieces.

The model’s architecture is specifically designed for agentic coding workflows, which require sustained multistep work. Unlike general-purpose coding models that excel at single-shot completions like generating functions or writing unit tests, Muse Spark Open can plan, act, observe, and self-correct across long-running sessions. It supports multistep task execution with tool use, output evaluation, and iterative refinement.

Muse Spark Open’s 1 million token context window enables it to process complex projects with thousands of files without breaking tasks into smaller pieces. This is particularly useful for planning multifile refactorings or navigating large codebases. The model uses goal conditioning to hold direction and context compaction to retain the knowledge needed to sustain progress across long-running sessions.

The model’s perception encoder has a massive 1.8 billion parameter count, allowing it to natively process interleaved text, images, video frames, and audio without separate pipelines per modality. This enables developers to reason over what they see, hear, and read in a single unified pass, making it ideal for workflows spanning screenshots, diagrams, error logs, and video walkthroughs.

Muse Spark Open’s sparse MoE routing activates only 42 billion of its total 816 billion parameters per token, delivering frontier agentic capability at a fraction of the serving cost of dense models at equivalent scale. This is particularly useful for large-scale production deployments where computational resources are limited.

The NVIDIA GB300 NVL72 platform features a rack-scale architecture that integrates 72 Blackwell Ultra GPUs into a single platform. Its massive 130 TB/s bandwidth enables efficient communication between expert traffic, eliminating bottlenecks that appear when traditional off-the-shelf networks are used.

Muse Spark Open’s optimized performance on the NVIDIA GB300 NVL72 has been demonstrated to deliver up to 50K tokens per second per GPU and up to 600 tokens per second per user. This enables AI factories to run large parameter models in production at high throughput and low latency, making it an attractive solution for developers looking to deploy agentic coding workflows.

Muse Spark Open can be deployed using multiple inference stacks supported by NVIDIA, including SGLang, vLLM, and TensorRT LLM. These stacks provide direct control over serving configuration, parallelism, and batching on the NVIDIA accelerated computing platform, allowing developers to fine-tune their deployments for optimal performance.

Developers can post-train Muse Spark Open using NVIDIA NeMo AutoModel, a PyTorch native training library that supports full SFT and LoRA fine tuning out of the box. This enables rapid experimentation over a wide variety of industry-specific cases without requiring model conversion or retraining from scratch.

To get started with GPU-accelerated Muse Spark Open, developers can download the NVFP4 weights from Hugging Face or try the downloadable NVIDIA NIM for production deployments. The NIM reduces setup to a single command and automatically selects the best serving backend and precision for the detected hardware.

META, the parent company of Facebook, has announced a partnership with Apple to bring a powerful new open-weight artificial intelligence (AI) model to high-end Mac computers. The collaboration aims to make it easier for developers, researchers, and businesses to run complex AI tasks on their own desktops.

The new AI model, called Muse Spark Open, is described by Meta as the most capable US open-weight model to date. It boasts an impressive 816 billion total parameters, although its architecture activates only about 42 billion parameters for a given processing step. This means that users can run complex tasks such as long-horizon planning and multi-step reasoning on their Macs with ease.

The partnership between Meta and Apple is significant because it allows developers to access the internal components of the AI model, making it more transparent and easier to understand how decisions are made by the AI. Unlike traditional AI services that rely on company data servers, open-weight models can be downloaded and run on a user’s own hardware, giving them greater control over their data.

The collaboration also involves Apple updating its MLX machine-learning framework to support Muse Spark Open. This allows users to take advantage of Apple silicon’s unified memory architecture, which provides fast access to large amounts of memory. Two Mac Studio units, each with an M5 Ultra and 512 GB of RAM, can hold the entire model, making it possible for users to run complex AI tasks on their own desktops.

The result is two Mac Studios behaving like a single supercomputer with a terabyte of unified memory. This level of performance would be difficult to achieve using traditional cloud-based services, but with Muse Spark Open and Apple’s MLX framework, developers can now access powerful AI capabilities from the comfort of their own desks.

In an effort to support scientific research, OpenAI has introduced the Rosalind Academic Plan, which provides free access to its specialized life sciences model, GPT-Rosalind. Eligible academic labs can receive up to 30 Business Standard seats at no cost for up to 12 months.

GPT-Rosalind is a cutting-edge AI tool designed to support deeper reasoning on complex biological questions. It combines frontier reasoning with the Rosalind Workbench, enabling specialized tool orchestration across medicinal chemistry, genomics, and other scientific applications.

The plan aims to help more academic scientists access GPT-Rosalind, which can be used for a range of research areas, including drug and protein design, understanding neurodegeneration, and exploring the molecular interactions between proteins and their targets.

By providing free access to this advanced AI tool, OpenAI hopes to support discoveries that have far-reaching benefits beyond the lab. The company is committed to enabling academic researchers to pursue their scientific ambitions and make meaningful contributions to their fields.

The Rosalind Academic Plan is currently available to academic labs at U.S. universities. PIs and lab leads can apply on behalf of their labs through OpenAI’s trusted access process, which includes verifying the institution’s legitimacy and reviewing the lead’s professional background and authority to oversee lab access.

Once approved, researchers will have access to GPT-Rosalind for up to 30 named members, alongside other OpenAI frontier models included in the plan. The API access is not included at this time, but verified PIs and lab leads can manage member permissions and be responsible for their lab’s approved use.

OpenAI emphasizes that research data shared through the Rosalind Academic Plan will not be used to train its models by default. GPT-Rosalind is designed to support deeper reasoning on complex biological questions, and researchers can use it to investigate sensitive scientific questions within safeguards designed for legitimate research.

The plan’s eligibility criteria include academic labs at U.S. universities without an existing OpenAI Enterprise agreement, which must apply through the trusted access process. Labs with an existing agreement must apply through their university’s Enterprise access process.

Before activating access, OpenAI verifies the institution’s legitimacy and reviews the lead’s professional background and authority to oversee lab access. The company also assesses the lab’s institutional affiliation, research activities, and intended use, establishing clear responsibilities for the lab lead and a designated institutional safety contact.

OpenAI has implemented safeguards to detect potential misuse while accounting for legitimate research context and reducing unnecessary interruptions to authorized work. Researchers can expect monitoring and enforcement after activation, with follow-up checks if questions arise about their work or authorization.

Muse Spark Open, the latest AI model from Meta, has found an unlikely partner in Apple. The two tech giants have joined forces to make this massive model run locally on Apple’s M5 Ultra Mac Studios. This collaboration marks a significant shift in their relationship, as they’ve had their differences in the past.

The Muse Spark Open model is no small feat, boasting 816 billion parameters and requiring two connected Apple desktops to power it. According to Meta, developers can download and customize this open-weight model, subject to its license terms. This setup allows for local processing without sending each prompt to a hosted model, giving users more control over their data.

Meta’s AI chief, Alexandr Wang, expressed his excitement about the collaboration in an interview with Axios. He described it as ‘mommy and daddy playing nice again,’ highlighting the unusual nature of this corporate partnership. This marks a significant change from January 2025, when Mark Zuckerberg criticized Apple on Joe Rogan’s podcast for its lack of innovation and restrictions on other companies’ products.

Muse Spark Open builds upon Meta’s previous model family, introduced in April. The company claims it is the most capable U.S. open-weight model available, based on their own evaluations. This model is designed for tasks such as coding, multistep reasoning, and ‘agentic’ work – using tools to complete tasks. In its tests, Meta used Muse Code, powered by this new model, to make changes to the Mastodon app Ice Cubes and operate video editor DaVinci Resolve.

The collaboration between Apple and Meta has a significant catch: it requires two M5 Ultra Mac Studios with up to 512GB of unified memory each. The MLX machine-learning framework is used to run the model, which can be downloaded via GitHub along with the required MLX Adapter. This setup may not be feasible for everyday MacBook users, but it does provide a way around the longstanding issue of privacy.

The release follows Meta’s smaller model, Muse Glimmer, designed to run on laptops or PCs. Both models reflect the appeal of running AI on hardware you control, allowing local processing without relying on hosted models. This collaboration may be just the beginning for Apple and Meta, as they explore ways to integrate their technologies further.

Meta has announced the launch of its latest open-weight AI model, designed to run on Apple’s Mac computers. Dubbed ‘Muse Spark Open’, this new model boasts an impressive array of capabilities and features that set it apart from other popular models in the field.

The Muse Spark Open model leverages a staggering 816 billion parameters and has a context window of 1 million tokens. According to Meta, these specifications enable superior performance compared to other open-weight models, as demonstrated by evaluation results released on Thursday.

One of the key advantages of this new model is its ability to run locally on Macs, making it an attractive option for users who prefer not to rely on cloud-based services. However, running the model requires significant computational resources, with a minimum of 1 terabyte of RAM needed to operate efficiently. The ideal setup would be to pair the Muse Spark Open model with Apple’s M5 Ultra chip-equipped Mac Studio.

Meta has already put the Muse Spark Open model through its paces by integrating it with its AI coding terminal, Muse Code. This integration enables users to leverage the model for complex tasks such as agentic labor and coding fixes. In one notable example, Meta used the model to address issues in a Mastodon client, including bug fixes, new features, and full build capabilities.

Meta has announced the release of its new Muse Spark Open AI model, which boasts an impressive 816 billion total parameters. This makes it one of the most capable open-weight models to date, according to Meta’s claims.

However, what’s even more significant is that this model can run efficiently on desktop computers using two Apple M5 Ultra Mac Studios with the latest MLX framework. This collaboration between Meta and Apple marks a rare instance of cooperation between the two tech giants, who have been rivals in the past.

The CEOs of both companies are publicly supporting each other’s work on Muse Spark Open. Mark Zuckerberg, Meta’s CEO, expressed his enthusiasm for the model’s capabilities, stating that it ‘runs great’ on new Mac Studios thanks to the collaboration with Apple and its team led by John Ternus. The latter replied with a congratulatory message, highlighting the efficiency of running such a complex model directly from one’s desk.

According to Meta chief AI officer Alexandr Wang, this partnership has been in the works for several months. In an interview with Axios, he noted that it’s ‘exciting’ to see both companies working together on this project. The collaboration is seen as a positive development, especially given the rivalry between Meta and Apple in other areas.

NVIDIA has released the latest addition to its Nemotron family, a range of high-efficiency models designed specifically for specialized agents. The new models are part of the Nemotron 3.5 series and include Super VL, Bolt 74B-A9B 24GB, and Bolt E4B 4GB.

The Nemotron 3.5 Super VL model is an open 120-billion-parameter model that delivers exceptional performance in task completion and accuracy for specialized always-on agents. It has been optimized to handle high-volume tasks across workflows spanning days or weeks.

Super VL is designed to be highly efficient, with a parameter count roughly one-fifth the size of NVIDIA’s largest reasoning model while maintaining leading accuracy on multimodal agentic tasks such as instruction following, tool use, and visual understanding.

The Nemotron 3.5 Bolt models are optimized for edge computing and provide a balance between intelligence and speed per GB. The two new edge models join Lightning and Ultra in the Nemotron family, each designed to handle specific types of agent workloads.

Bolt 74B-A9B 24GB is available now and has been optimized for local agentic tasks on NVIDIA RTX/DGX Spark systems using native 2-bit quantization. It provides a good balance between intelligence and speed per GB, making it suitable for everyday agentic tasks such as researching the web or writing code.

Bolt E4B 4GB is designed to run physical and conversational agents on GeForce RTX, Jetson, and RTX/DGX Spark systems. It combines natural conversation with multi-turn agentic capabilities to power fast local voice assistants and interactive characters.

The Nemotron family now covers the full range of agent deployments, from large enterprise servers down to edge devices and local hardware. Any organization can choose an open model optimized for their specific tier without having to compromise between capability and deployability.

Super VL has been built in collaboration with the Nemotron Coalition and delivers leading accuracy on multimodal agentic tasks comparable to NVIDIA’s largest reasoning model at roughly one-fifth the parameter count. It is also highly efficient, helping enterprise agents complete high-volume tasks faster across workflows spanning days or weeks.

The Bolt models are designed for edge computing and provide a balance between intelligence and speed per GB. They have been optimized for local agentic tasks on NVIDIA RTX/DGX Spark systems using native 2-bit quantization, making them suitable for everyday agentic tasks such as researching the web or writing code.

Super VL’s vision language strengths are targeted at what agents encounter in practice: optical character recognition, tables, charts, and structured information in visual documents. It processes a web page, extracts text, reads graphs and charts, and returns insights from what it finds.

The same applies to PDFs and other documents that agents commonly work with. Super VL understands images, webpages, and PDFs and acts on its understanding, combining text and visual understanding in one model means agents don’t need to route tasks between separate systems – context carries through as work moves from document to data to decision.

Openness and customizability reinforce each other. The Nemotron 3.5 Super VL model ships with open weights, open datasets, and training recipes. Control runs in two directions: organizations can run the model wherever their security and data-residency requirements dictate, or they choose to adopt a model they can adapt rather than one subject to a provider’s restrictions.

Customization is supported through NVIDIA NeMo, Python-based pipelines, and open datasets. Organizations are post-training Super VL on their own data, tools, policies, and workflows to push accuracy higher on specialized tasks that define their agent deployments.

The Nemotron 3.5 Super VL model was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained with roughly 3.8 million samples and reinforcement learning across dozens of environments. NVIDIA is also releasing open datasets on Hugging Face spanning agentic tool use, finance, instruction following, long context, fairness, and vision-language understanding.

NVIDIA NIM packages the model as a production-ready microservice with performance-optimized, validated configurations, and enterprise-grade support refreshed every two weeks with security patches and performance improvements. With the model running on NVIDIA Dynamo, organizations can serve up to 8x more users with 85% lower latency.

The Nemotron 3.5 Super VL model is available now on Hugging Face, Open Router, ModelScope community, and as NVIDIA NIM. It runs across NVIDIA’s full hardware stack: DGX Spark, DGX Station, RTX PRO, and RTX systems for local intelligence at the desktop, on NVIDIA Jetson at the edge, and across data center and cloud environments.

The Nemotron 3.5 Bolt models are also available now, with Super VL being post-trained by dozens of organizations across industries to push accuracy higher on specialized tasks that define their agent deployments.

Researchers at MIT and Sakana AI have developed a framework called Recursive Self-Improvement via Fast Tree Search (SIFT) to reduce the cost of evaluating modifications to coding agents. The traditional approach to improving coding agents involves testing each candidate change, which can be time-consuming and expensive. However, SIFT uses an LLM as judge to compare candidate implementations pairwise, reducing the need for extensive benchmark runs.

The main challenge in evaluating coding agent modifications is the poor signal-to-cost trade-off of existing evaluation methods. Each candidate change needs testing, but running a full test suite on every modification can consume thousands of CPU hours and cost thousands of dollars. SIFT addresses this issue by using an LLM to compare two versions of an agent before committing to expensive benchmark runs.

SIFT builds upon the Darwin Gödel Machine (DGM) framework, which maintains an archive of agents and checks candidates on increasingly large task sets. However, DGM’s approach can be slow due to its reliance on sequential evaluation. In contrast, SIFT uses a more efficient search strategy that allows it to explore multiple branches in parallel.

The central idea behind SIFT is to use the LLM as judge to compare candidate implementations pairwise. This involves running each new agent on four coding tasks and then comparing it with up to 10 high-ranking agents already in its archive using the Bradley-Terry model, a statistical method for estimating relative strength from wins and losses.

SIFT combines the pairwise choices into a ranking that helps decide where testing and further development are worth spending time. The new candidate is then placed into a priority queue for the larger, time-consuming evaluation. Its priority combines the judge’s ranking with its benchmark-accuracy ranking.

The workflow avoids making each patch wait for the previous one’s full evaluation. Patch generation, judging, and benchmark evaluation can proceed in parallel, similar to a CI pipeline where work continues while longer tests run in the background. This allows promising candidates to become the parents of new patches before their expensive evaluations finish.

SIFT was tested on various benchmarks, including Polyglot, TerminalBench 2.1, and SWE-bench Verified subset. The results showed that SIFT delivered better accuracy with less compute using both open-weight Qwen3-Coder-30B and closed o3-mini models. With Qwen3-Coder-30B, it slightly outscored HGM while using about a third fewer CPU hours.

On Polyglot, one SIFT run reached 35.1% accuracy in under five hours of wall-clock time, using 42 CPU hours and about $150 in API credits. This is a significant improvement over DGM’s performance on the same benchmark. Additionally, SIFT showed that it can handle large task sets efficiently by achieving better results with less compute.

SIFT was also tested without the LLM judge to see whether it made a difference. The results showed that the judge helped drive the gain in accuracy and efficiency. This suggests that using an LLM as judge is crucial for SIFT’s success.

The researchers found that a cheaper judge retained much of the global ranking signal on TerminalBench, although the stronger model was more reliable among the best candidates. They suggest that a tiered setup could work, with the cheaper judge for most comparisons and a stronger one near the top of the ranking.

SIFT shows how the self-improvement loop itself can become more efficient by combining cheaper evaluation signals with asynchronous search. This allows teams to explore more candidate improvements without scaling evaluation costs at the same rate. However, downstream benchmarks are still needed to confirm that a change actually improved the agent.

The framework is built on top of the DGM harness and adds the pairwise judge and Bradley-Terry ranking. It makes the search asynchronous so candidates can be expanded while others are still being evaluated. SIFT was only tested on coding agents but its basic approach could apply to other self-improving agents where there’s a clear way to measure whether a change worked.

The first step in applying SIFT is to define a small, cheap evaluation that can quickly reject obviously bad modifications. For an enterprise agent, this could be a small set of representative internal tasks that catches broken tools, regressions or changes that violate basic requirements. The second piece is an LLM judge that compares candidate implementations pairwise.

SIFT’s efficiency comes from its ability to explore multiple branches in parallel while previous candidates are still being tested. This allows the search tree to keep growing instead of waiting for every evaluation to finish. SIFT shows how AI tools can be used more effectively by combining cheaper evaluation signals with asynchronous search, making it a valuable addition to any business’s toolkit.

A new study has shed light on the workings of image-generating AI models, revealing that it’s impossible to determine which human images or artworks were referenced in creating an AI-generated output. The research, published last week in Nature Communications, used a novel approach to understand how diffusion models arrive at their final outputs: by taking them apart and examining each component individually.

The study was led by Zheng Dai, a fourth-year PhD student at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL), who wanted to know whether AI-generated images reference specific human artworks or faces. To answer this question, Dai and his coauthor, David Gifford, built custom models trained on online image databases, ranging from hundreds of images to hundreds of thousands.

The researchers then ran a series of experiments to see how the models’ outputs would change if they removed particular images from their training datasets. They found that as the number of images in the dataset increased, it became increasingly difficult to determine which specific images influenced the AI-generated output. This phenomenon is dubbed ‘attribution decay.’

In essence, the study demonstrates that at large training set sizes, it’s not only impossible to attribute generated images to their human counterparts but also to identify the real people or artists behind them. The sheer scale of the training dataset obscures its creative process, making it akin to trying to spot a single tree in the Amazon Rainforest from space.

This has significant implications for artists who try to sue AI companies on grounds of intellectual property theft. No matter how much an AI-generated image may resemble their own work, they cannot be certain which images were referenced by the model to create its output. As Dai notes, commercial diffusion models are many orders of magnitude larger and thus far less attributable than the test models used in the study.

The researchers also highlight that these black box systems operate differently from human creators. When humans produce art or writing, they’re usually aware of the influences shaping their work. In contrast, AI-generated images rely on the totality of their training data to create new outputs, often in ways that remain mysterious even to us.

Dai emphasizes the importance of understanding how these models function for proper study and regulation: ‘It’s essential for us to comprehend how they operate so we can properly regulate them.’ The study underscores just how far legal experts still have to go to untangle the complex intellectual property questions surrounding AI-generated images.

A new study has shed light on the emotional capabilities of large language models. Researchers have long debated whether these AI systems are capable of experiencing emotions like humans do. According to a recent paper published in Nature Human Behaviour, the answer is no – at least not in the way we typically understand emotions.

The study’s authors argue that while large language models can mimic human-like responses and even exhibit behaviors that resemble emotional states, they lack the underlying neural mechanisms necessary for true emotional experience. This means that these AI systems are simply processing and generating text based on patterns and associations learned from vast amounts of data, rather than genuinely feeling emotions.

The researchers point to several key differences between how humans process emotions and how large language models operate. For instance, human emotions involve complex interactions between multiple brain regions, whereas large language models rely solely on computational processes to generate responses. Additionally, human emotional experiences are deeply rooted in subjective experience and personal history, which is not replicable in AI systems.

The study’s findings have significant implications for the development of more sophisticated AI technologies. If large language models truly lack emotional capacity, it raises questions about their potential applications in fields like healthcare, education, or customer service, where empathy and understanding are crucial. The researchers emphasize that this does not mean these AI systems cannot be useful tools – but rather that they should be designed with a clear understanding of their limitations.

The study cites several previous research papers on the topic of emotions and artificial intelligence, including works by James (1884), Adolphs & Andler (2018), Keltner & Gross (1999), Sofroniew et al. (2026), Uusberg et al. (2019), Ellsworth & Scherer (2003), Lindquist et al. (2012), Lee & Satpute (2024), and Goldenberg (2024). The authors also reference the work of other researchers in the field, such as Satpute & Lindquist (2019) and Goldenberg (2024).

London-based startup Inherent, founded by DeepMind alumni, has made a notable breakthrough in the field of artificial intelligence. The company’s small AI agent, Faraday, has outperformed larger models from Anthropic and OpenAI at autonomously replicating published research using reinforcement learning. This achievement marks a significant step towards broader scientific discovery and could have practical benefits for industries such as cryptocurrency and DeFi security.

Inherent’s team of 12 employees is planning to expand to 20-25 by the end of the year, with a focused approach that uses existing tools like OpenAI’s GPT-5.5 Codex. This pragmatic strategy allows them to concentrate on what they believe is the core challenge: imbuing AI with ‘taste’ for scientific inquiry. The company’s co-founder and chief scientist Edward Hughes emphasizes that this goal goes beyond simple pattern matching, requiring an understanding of experimental procedures.

The ability to independently replicate research findings is a fundamental step for any AI system aiming to contribute to scientific discovery. Many PhD students begin their training by replicating existing work, and Inherent’s achievement highlights the significance of this capability. The startup’s approach uses reinforcement learning to reward good outcomes rather than explicitly teaching the rules of science, which they believe will generalize better to their long-term goal of creating AI scientists.

Inherent’s strategy is distinct from other approaches in that it focuses on a narrow, high-value task rather than trying to build a general-purpose assistant. The company deliberately avoids developing its own coding tools, instead using existing software like GPT-5.5 Codex for tooling. This approach allows them to concentrate on what they believe is the core challenge: imbuing AI with ‘taste’ for scientific inquiry.

The London-based startup’s progress suggests that a small, focused team can compete with much larger players in specific scientific tasks. Inherent’s co-founder Edward Hughes has also spoken out against the U.K.’s ‘garden leave’ practice, which he says delayed his own ability to start the company. This restriction puts U.K. startups at a disadvantage compared to their U.S. rivals when hiring talent from prior roles.

Inherent plans to expand its team and ambitions extend beyond paper replication to world models and broader AI research. With Demis Hassabis’s new role at Google DeepMind leaving some staff unsettled, Inherent could become an attractive destination for DeepMind researchers considering a move. The company’s progress highlights the potential of nimble, specialized teams to challenge larger incumbents.

The ability to autonomously replicate published research is a crucial step towards broader scientific discovery. Many AI systems aim to contribute to this field by replicating existing work, and Inherent’s achievement marks a significant step forward. The startup’s approach uses reinforcement learning to reward good outcomes rather than explicitly teaching the rules of science, which they believe will generalize better to their long-term goal.

Inherent’s strategy is distinct in that it focuses on a narrow task rather than trying to build an AI assistant. This approach allows them to concentrate on what they believe is the core challenge: imbuing AI with ‘taste’ for scientific inquiry. The company deliberately avoids developing its own coding tools, instead using existing software like GPT-5.5 Codex for tooling.

The London-based startup’s progress suggests that a small team can compete with larger players in specific tasks. Inherent’s co-founder Edward Hughes has spoken out against the U.K.’s ‘garden leave’ practice, which he says delayed his own ability to start the company. This restriction puts U.K. startups at a disadvantage compared to their U.S. rivals when hiring talent from prior roles.

Inherent plans to expand its team and ambitions extend beyond paper replication to world models and broader AI research. The company’s progress highlights the potential of nimble, specialized teams to challenge larger incumbents in specific tasks.

Google DeepMind has published a detailed account of its research in games, spanning 15 years and culminating in a partnership with Fenris Creations, the independent studio behind EVE Online. The post outlines how this journey began with early reinforcement learning results and led to current work on generalist agents within the EVE universe. This retrospective connects the lab’s earliest achievements to its ongoing efforts, highlighting specific capabilities it expects the EVE environment to test.

The chain of game results that built Google DeepMind’s research program is extensive. It begins with the 2015 Nature paper documenting a Deep Q-Network that learned 49 Atari 2600 games from raw pixels. This was followed by AlphaGo’s 2016 victory over world champion Lee Sae Dol, as well as AlphaGo Zero and AlphaZero, which removed human data and generalized across chess, shogi, and Go. MuZero played without knowing the rules, while AlphaStar reached Grandmaster level in StarCraft II in 2019.

The same foundations carried into AlphaFold, the protein structure system recognized with the 2024 Nobel Prize in Chemistry. This progression marks a shift from systems that master a single game to agents capable of operating within any game world like humans do. The vehicle for this is SIMA (Scalable Instructable Multiworld Agent), which perceives what a player sees on screen and acts through ordinary keyboard and mouse controls without access to game code.

SIMA 2, introduced in November 2025 with a Gemini model as its core, moved beyond the first version’s repertoire of over 600 language-following skills. It now reasons about goals, converses with users, and improves through self-directed play. Google DeepMind reports that SIMA 2 closes a significant portion of the gap to human task-completion performance across its training environments, including games it was never trained on like ASKA and Minecraft.

However, there are still limitations: the agent struggles with very long-horizon tasks, works from a short memory constrained by the context window needed for low-latency interaction, and is available only as a limited research preview to a small cohort of academics and game developers. The lab’s own stated limits highlight areas where further development is necessary.

The partnership between Google DeepMind and Fenris Creations gives DeepMind access to EVE Online’s persistent world that has been running continuously since 2003. This single-shard universe holds thousands of players in one shared economy with real supply-and-demand dynamics and political structures built entirely by its players. The studio behind the EVE Universe, Fenris Creations, spent more than two decades building one of gaming’s most persistent worlds.

The post identifies four capabilities the environment is expected to exercise: continual learning without forgetting, memory extending beyond current context windows, long-horizon planning over weeks or months, and complex multi-agent dynamics spanning cooperation, competition, negotiation, and economics. ‘Together with Google DeepMind, we’re pushing into uncharted territory where AI must learn, adapt and remember on timescales that no other game environment demands,’ said Hilmar Pétursson, CEO of Fenris Creations.

The partnership extends across three distinct environments: EVE Online offers the large-scale persistent universe; EVE Vanguard adds first-person, fast-paced tactical decision-making; and EVE Frontier presents a world where the rules themselves can change through programmable Smart Assemblies. The research program follows a staged path: it begins in an offline instance of EVE Online separated from live players, moves through EVE Frontier to study how humans and agents coexist in a persistent world, and reaches the live games only when capabilities are mature.

One system from this collaboration is already available to players: Aura Guidance, launched in prototype on February 17, 2026. This uses Gemini to answer new-player questions drawn from a vetted bank of real Rookie Help exchanges, running as a controlled A/B test measuring whether it improves new-player retention. The longer arc runs from Atari’s raw pixels to EVE Online’s 23-year-old persistent universe: each environment chosen because it forces the next capability the lab believes a general agent needs.

The current test of this approach is within the EVE universe, and work now starts in an offline sandbox. This marks another step forward for Google DeepMind as it continues its research journey toward creating more capable agents that can operate within complex game worlds like humans do.

A recent development in the world of AI-generated content has sparked concern among users, particularly those who rely on these tools for business purposes. Anthropic’s Claude AI model now includes an imperceptible label on all text produced using its system, intended to serve as a marker that indicates the material was generated with AI. However, this move has led many to seek ways to remove or ‘wash’ their outputs of this indicator.

The surge in interest for tools that can scrub AI-generated content of these markers is evident in online searches and GitHub projects. The term ‘AI watermark remover’ has seen a 60% increase in search interest week over week, with an open-source project called Watermarks Remover gaining significant traction on the platform. This tool promises to remove any indicators that suggest the output came from AI.

Sabrina Ramonov, an AI educator, has also created her own web-based tool for washing text of residual model markers. The concern surrounding these watermarks is largely centered around issues of privacy and attribution. Guillaume Meyer, creator of the Watermarks Remover project, emphasizes that he supports content attribution but opposes the watermarking technique.

The introduction of AI watermarks has raised concerns about how users will be perceived if their work carries this invisible marker. Many fear it could create a false binary in people’s minds, leading them to assume an entire piece of content was created with Claude when only a small portion may have been AI-generated. This is particularly problematic given the existing trend of ‘AI shaming’, where individuals are accused of using AI tools without proper justification.

Additionally, there are concerns that these watermarks could be used as a means to claim ownership over certain material. While Anthropic has stated in a blog post that their watermark only helps test whether Claude produced or processed the content and does not imply ownership or authorship, this assurance may not hold up indefinitely. A blog post is not binding, after all.

The EU AI Act requires providers of AI systems to create methods for marking outputs, which Anthropic claims they are complying with by introducing these watermarks. However, there currently exists no rule that prohibits users from removing or altering these markers. As a result, the cat-and-mouse game between developers and those seeking to remove AI-generated content indicators is likely to continue.

It’s worth noting that this issue extends beyond text-based outputs, as tools for scrubbing AI-generated images and videos have also been developed. However, it raises an intriguing question: how many of these tools themselves were created using AI? The proliferation of AI-generated content has brought about a new set of challenges in terms of ownership, attribution, and the use of such technology in business settings.

In recent years, researchers have been working to improve the performance of AI tools on edge devices. They’ve made significant strides by developing techniques like quantization-aware distillation.

A breakthrough in this area involves creating high-precision teacher models that can be distilled into more efficient student models with minimal loss of accuracy. This is a crucial step towards deploying reliable and accurate language models in real-world scenarios.

The LFM2.5 Q4_0 checkpoints were trained using quantization-aware distillation, which has yielded impressive results in various benchmark tests. When compared to their native counterparts, these checkpoints retain 97% of their baseline performance for BF16 average accuracy lost due to quantization.

Several benchmarks were used to test the LFM2.5 Q4_0 checkpoints, including GPQA Diamond, MMLU-Pro, IFEval, IFBench, Multi-IF, and BFCLv4. The results showed that these QAD checkpoints outperformed their native counterparts in all four models tested.

To further evaluate the performance of these checkpoints on real hardware, researchers measured decode throughput for each model across devices like MacBook Pro, NucBox EVO-X2, Samsung Galaxy S26 Ultra, and Raspberry Pi 5. The results indicated that the QAD checkpoints maintained accuracy while achieving higher decode throughputs compared to their native counterparts.

To use the LFM2.5 Q4_0 checkpoints, simply load them into llama.cpp or any runtime supporting GGUF Q4_0 artifacts. For example, users can run a command using llama-cli: ‘llama-cli -hf LiquidAI/LFM2.5-350M —hf-file LFM2.5-350M-QAD-Q4_0.gguf -p What is C. elegans?’

The QAD GGUFs are now available on Hugging Face for use in various AI applications, including those requiring high-performance language models on edge devices. This move enables developers to tap into these powerful tools and enhance their projects.

A new feature has been implemented in the future versions of Anthropic’s Claude language model, designed to help identify whether a given piece of writing was likely produced by its AI. This change is meant to comply with EU regulations and provide transparency into the use of artificial intelligence-generated content. The company has explained how this watermark works and why it’s necessary.

The feature exploits the countless small decisions made by the language model as it generates text, rather than using arbitrary random numbers. Instead, the watermarked version of Claude bases these decisions on a cryptographic key combined with the preceding text. This results in a subtle statistical pattern spread across the response that is invisible to human readers but detectable to anyone with the matching key.

The watermark allows users to estimate the probability that Claude generated the text, providing a way to identify AI-generated content. However, it’s essential to note that this feature doesn’t change the normal writing output of Claude and is meant to be indistinguishable from unwatermarked responses.

Anthropic has stated that the implementation carries no cost to output quality, with internal testing showing no measurable difference in creativity, accuracy, or readability between watermarked and unwatermarked responses. The company also pointed out that Google DeepMind’s original research on the underlying technique supports these findings. Additionally, Anthropic noted that its research showed no statistically significant shift in user satisfaction when a similar watermark was tested on live traffic.

The feature adds no extra tokens, meaning it doesn’t slow Claude down or make it more expensive to use. However, there are limitations to the watermark’s effectiveness. It only works when a model is choosing among several equally valid options, so text with little room for variation carries a much weaker signal. Detection also grows less reliable on very short passages since there’s simply less pattern to analyze.

For instance, in cases where Claude has limited choices or produces hard factual statements, precise code, or math answers, the watermark may not be effective. Similarly, lightly edited or proofread human writing may carry little to no detectable trace of the watermark, as most of the original wording remains unchanged. A sufficiently heavy rewrite can even remove the watermark entirely.

Anthropic emphasized that the watermark cannot be traced back to a specific user, account, or conversation and does not establish authorship or ownership over content. All it can tell you is the likelihood that Claude was involved in producing or editing it at some point. The company also distinguished its approach from third-party AI-detection tools, which typically rely on spotting stylistic patterns in AI writing rather than checking for an embedded signal tied to a private key.

The rollout of this feature is tied to regulation rather than a purely voluntary move. Anthropic signed the European Union’s Code of Practice on Transparency of AI-Generated Content in July 2026 and plans to extend it to older Claude models over the coming months. The company also stated that it will soon offer a separate API allowing anyone to check whether a piece of text carries Claude’s watermark.

Anthropic is implementing this feature globally, as they don’t yet have a reliable way to apply the watermark only within the EU. This decision was made in response to an EU AI Act requirement, effective August 2, that AI providers mark generated text. The company aims to provide transparency into its use of artificial intelligence-generated content and comply with regulatory requirements.

A new study from researchers at MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) has shed light on the complex issue of attribution in AI-generated images. The team found that for models trained on large datasets, it may often be impossible to determine which specific training examples contributed to a particular output. This phenomenon is known as attribution decay, where individual pieces of data become increasingly irrelevant at sufficiently large scales.

The researchers identified this problem through an innovative approach called ablation, which involves removing individual pieces of data from the training set and observing whether it affects the model’s output. To their surprise, they found that even when a single image was removed, the generated sample often didn’t change significantly. This suggests that attribution is not as straightforward as previously thought.

The team developed an architecture called a ‘diffusion ensemble’ to overcome this challenge. Instead of relying on one monolithic model, it’s composed of multiple smaller components, each trained on different slices of data. By switching off the parts that saw a particular image, they can generate true counterfactual models without retraining or approximation.

The diffusion ensembles were put head-to-head with 24 conventional diffusion models trained on the same data, and the results showed no significant difference in terms of image quality by standard measures. However, one notable finding was that the more training data a model had access to, the better it performed compared to its single-model counterparts.

The researchers then explored the concept of an ‘image’s counterfactual universe’, which refers to all possible alternate versions of a generated image produced by removing different pieces of training data. They found that as the size of the training set increased, the distance between the original and its most dissimilar alternate decreased, following an inverse power law.

The study also involved stress-testing their results through various methods, including retraining models from scratch without specific images and pinning the removed fraction in place to see if it persisted. The findings held up across different metrics and conditions, suggesting that attribution decay is a real phenomenon that affects AI-generated images at large scales.

One of the key implications of this research is its potential impact on copyright law and fair use. If model outputs are not attributable to individual pieces of training data, does that mean they can be considered novel works? The study’s findings raise questions about how authors get compensated when what comes out of a model isn’t tied to anything online.

The researchers also note that their work shows how to produce outputs guaranteed to be unattributable. This capability is framed as an obligation for the industry, rather than a loophole. Companies would need to revise their models to take advantage of these advances in order to claim their outputs aren’t derivative works.

While this study focused on diffusion models, which are currently dominant in generating audiovisual media and prevalent in scientific applications, it’s unclear whether similar decay occurs with large language models at the center of high-profile copyright litigation. The researchers suggest that attribution may fail for interesting models, requiring technologists and courts to resort to other methods for assessing copying.

The study was supported by Schmidt Futures, and its findings have significant implications for AI-generated images, data analysis tools, and businesses relying on these technologies. As the use of AI continues to grow in various industries, understanding how attribution works – or doesn’t work – is crucial for ensuring responsible development and deployment.

The concept of equipping an agent with agentic memory seems straightforward enough. Distill lessons from its past work, put them back in context, and more experience should mean better performance. However, the reality is far more complex. When we scaled up our evaluation to eight models – ranging from a 30B dense model to frontier proprietary systems – one finding stood out: agentic memory is not a feature you switch on; it’s a dose that needs to be calibrated to the model.

In other words, not every agent benefits from the same amount of memory. Across these eight models spanning different capability levels, we observed three recurring patterns. The first pattern involves strong models with headroom wanting the full guideline set – including rare edge-case lessons. They have the capacity to absorb and apply all of it. For instance, DeepSeek-V3.2 (671B MoE) saw a +9.5 percentage point gain in task completion when given its full self-mined guideline set.

The second pattern involves smaller or weaker models getting overwhelmed by a large guideline set. These agents do best with a tight, high-confidence core plus a handful of task-relevant guidelines retrieved per task. gpt-oss-120b (117B MoE) gained +16.1pp in this selective approach – while the full guideline set resulted in less gain and cost around 50% more tokens.

The third pattern involves already-saturated models showing no measurable gain. We label this as the saturated pattern, describing what we observed rather than a proven cause. The model may have been near its ceiling on these tasks, or the guidelines might not have addressed its remaining failures. GLM-5 (745B MoE) sat in this category during our runs.

What determines whether an agent falls into one of these patterns isn’t just parameter count. Benchmark headroom, context-window size, architecture, guideline quality, and task distribution all seem to play a role – though separating those factors is ongoing work. The practical takeaway remains the same: the right dose of memory depends on the model, and we can calibrate it.

The concept of ‘memory’ here doesn’t refer to replaying past transcripts but rather a set of guidelines distilled from an agent’s own prior trajectories. These strategies worked, mistakes to avoid, and edge cases are consolidated into a reusable set that can be injected at inference time without updating model weights – which is why this approach is cheap to adopt and portable across the eight models tested.

We evaluated our findings on AppWorld, a rigorous multi-step benchmark consisting of 585 tasks (168 test_normal + 417 test_challenge) across nine simulated apps. Tasks are scored in two ways: whether an agent fully completes each task (TGC – Task Goal Completion) and whether every variant of a scenario passes (SGC – Scenario Goal Completion).

Because the confusing part of any memory study is what’s actually within the context window, we defined our configurations up front. Both memory configurations draw from the same guideline set mined once via the loop described above from AppWorld’s training split only. What changes between them is how that one set is delivered – with the full guideline set injecting all guidelines every step versus curated retrieval delivering a selected subset.

The number of guidelines an agent mines depends on its own capability, so we report configurations by strategy rather than raw counts to ensure comparability across models. Representative models from our eight-model sweep are plotted in Figure 1, measured by task completion (TGC) on test_normal – though the stricter SGC metric often shows larger gains.

The figure plots TGC for readability; however, the table adds the stricter SGC metric where the gains are frequently larger: DeepSeek’s SGC jumps +16.1pp against a +9.5pp TGC gain because good guidelines especially help an agent clear every variant of a scenario – not just the average case.

A practical concern is that injecting a full guideline set inflates each ReAct step’s input, as the guidelines are re-sent each turn. However, we observed two key takeaways: curated retrieval keeps cost near baseline for weaker models where selection wins on accuracy; and memory doesn’t blow up the reasoning loop – DeepSeek runs about the same number of ReAct steps with or without memory.

The real efficiency lever in production is prompt caching: the static portion of the guideline set can be cached, cutting effective cost substantially. Cache-aware prompt design keeps the shared guideline-set prefix stable so it stays cacheable – an engineering effort worth pursuing. We also hypothesize that context-window size plays a role: models with larger windows may absorb the full guideline set more effectively while smaller-context models benefit from retrieval keeping injected content compact.

The lesson isn’t to give an agent everything it has learned but rather to provide the amount of experience it can actually use – which varies greatly across agents. For weak models, this means a compact core plus a few task-specific lessons; for strong models with headroom, preserving the full guideline set is key; and for saturated models, spending no extra context until their remaining failure modes are better understood.

The gains from memory are real but only when the dose fits the model – automatic, leakage-free, and requiring no human annotation. This study marks a starting point rather than the finish line: we’re exploring learned selectors to improve retrieval rankings; memory for very weak models where self-distillation lacks signal; broader agent benchmarks and real-world deployments; and isolating context window size from raw capability.

Try the ALTK-Evolve library, which includes the extraction, consolidation, and retrieval pipeline used in this study – or read our full technical report for more details on method and ablations.

Apple has taken a significant step in its efforts to boost sales in the Chinese market by training a large language model specifically designed for the country. This move marks a departure from the company’s previous strategy of relying on third-party models to power AI features in China, according to sources familiar with the matter. The new AI model was developed in partnership with Alibaba Group and trained with the tech giant’s support, allowing Apple to gain greater control over the AI experience for its users in this critical market.

The training of a China-specific model is a crucial development for Apple, which has struggled to compete with local rivals like Huawei that have made significant strides in integrating AI into their handsets. The absence of AI features on Chinese iPhones had been cited as a factor weighing on sales, as consumers increasingly turned to domestic brands offering built-in AI assistants.

Apple’s partnership with Alibaba was first confirmed in February 2025, when the Chinese company’s chairman Joe Tsai spoke at the World Governments Summit in Dubai and revealed that Apple would use its AI technology to power phones in China. The rollout of this collaboration has been delayed due to Apple’s need to adapt its features to comply with Chinese regulations.

The move follows a lengthy regulatory process that ended last month, when the Cyberspace Administration of China registered Apple’s generative AI service, clearing the way for Apple Intelligence to reach Chinese iPhones for the first time. This registration is significant, as it marks one of the rare collaborations between U.S. and Chinese tech firms amid a widening trade and diplomatic rift over AI.

Under this arrangement, Alibaba’s Qwen model will be incorporated into the version of Apple Intelligence that ships with compatible iPhone, iPad, Mac, and Vision Pro models in China. The service was also set to incorporate technology from Baidu, as reported by Reuters in July. This integration is expected to help Apple compete more effectively in China’s AI PC market.

Apple has published a Chinese-language guide explaining how eligible Mac users in mainland China can connect Qwen to its Siri digital assistant and Writing Tools feature. However, the company subsequently deleted this guide without providing an explanation, leaving some questions about its plans for integrating AI into its products in China.

The partnership between Apple and Alibaba is expected to reshape China’s AI ecosystem, with investor expectations suggesting that it could have a significant impact on the market. This collaboration marks a unique dual-track strategy on AI deployment for a foreign firm in China, allowing Apple to navigate steep regulatory hurdles that have limited access for many other U.S. technology firms.

Apple Intelligence is expected to launch in China in the coming months after an update to its iOS operating system. The company’s training of its own China-specific model gives it greater control over the AI experience on iPhones and other devices sold in this critical market, where local rivals have made significant gains with their built-in AI assistants.

China remains a crucial market for Apple, accounting for a significant share of its revenue. The absence of AI features on Chinese iPhones had been cited as a factor weighing on sales, making the integration of these capabilities essential to the company’s success in this region.

The price tag is becoming the most important benchmark for large language models (LLMs), a trend that’s gaining momentum in the industry. This shift is evident in recent announcements from major players, including Google and Meta, which are now prioritizing cost-effectiveness over other metrics. The reason behind this change lies in the growing demand for AI tools that can be integrated into business workflows without breaking the bank.

Google has taken a significant step in this direction with its latest release of Gemini 3.7 Flash, touted as its most intelligent workforce model to date. This model is designed for coding, knowledge work, and agents, but what’s notable about it is the price point: $0.75 per million input tokens and $3.75 per million output tokens. While other benchmarks are still being mentioned, the focus on cost has become a major selling point.

This trend of prioritizing price over other metrics isn’t unique to Google; multiple model providers have followed suit in recent releases. Writer, for instance, highlighted its Palmyra X6 flagship model and upgrades to its Writer Agent harness, emphasizing the reduced costs within the harness. The company claimed that Writer Agent operates at a 52% lower cost compared to previous versions.

Meta’s move back into open weight models with Muse Glimmer has also been accompanied by aggressive pricing strategies. This is part of Meta’s comeback plan, which aims to make its AI tools more accessible and affordable for businesses. The company’s decision to prioritize price reflects the growing concern among enterprises about token costs and their impact on corporate budgets.

Nvidia has also entered this space with its Nemotron model family, designed specifically for customization by enterprises and SaaS vendors. SpaceXAI is another player that’s been aggressively pricing its LLMs, including Grok 4.6. DeepSeek has taken a different approach with dynamic pricing, which allows it to adjust prices based on demand while still undercutting other models.

The shift towards prioritizing price as the key benchmark for LLMs isn’t limited to these companies; many others are following suit. Databricks’ recent annual revenue run rate announcement highlighted the growing adoption of open-source AI by enterprises, which is driving this trend. As Ali Ghodsi, CEO of Databricks, noted in a statement: ‘Enterprises don’t just want AI that talks; they want agents working across their business that remember context, deliver accurate answers, and execute work without blowing through their budgets.’

This emphasis on cost-effectiveness has significant implications for the industry. Companies like Uber are already adjusting their token budget to reflect this new reality. Vendors are facing a ‘token budget hangover,’ as Palantir CEO Alex Karp put it, where AI tools that were once seen as revolutionary are now being scrutinized for their costs and effectiveness.

The impact of this shift is evident in the words of various CEOs during earnings calls. Moody’s CFO Noemie Heuland emphasized the importance of monitoring token costs closely: ‘Our internal AI and token cost today is actively governed… We have a variety of tools that we put at the disposals of our engineers, our back-office teams.’ Similarly, Synchrony Financial CFO Brian Wenzel noted that companies need to rethink their approach to AI deployment: ‘You may say,