AI News Daily

Meta releases lightweight model Muse Ember, focused on cost-effectiveness for agentic tasks

Meta has released Muse Ember, its smallest and most efficient model, aimed at complex reasoning execution and everyday agentic work; it can carry workflows across applications, call tools in sequence, run tests and make targeted code fixes. The model supports a 1M-token context window and natively understands text, images, audio and video. On the Artificial Analysis Intelligence Index, its performance is comparable to Claude Haiku 5.5, at about one-third the per-task cost of the latter, and it scores higher than GPT-6 Luna under all reasoning settings. Muse Ember is now available on the Meta Model API, priced at $0.10 per million input tokens and $0.50 per million output tokens.

• Artificial Analysis evaluation: Meta Muse Ember scores 36 on the intelligence index, on par with Claude…Artificial Analysis
• Meta releases lightweight model Muse Ember, now available on the Meta Model APIAI at Meta
• Meta releases lightweight model Muse Ember, aimed at agentic executionTechCrunch AI
• Artificial Analysis evaluates Meta Muse Ember: intelligence index 36, cost at…Artificial Analysis

Meituan open-sources LongCat-2.5: a native multimodal model with 1.6T total parameters and about 48B activated

Meituan LongCat has released the weights of LongCat-2.5, with 1.6T total parameters and about 48B activated, natively multimodal. The model combines stronger visual understanding and agentic coding capabilities in a single model, covering everything from complex coding projects to creative design and 3D tasks. The weights are released under the MIT license and are now available on Hugging Face.

NVIDIA releases the GR00T 2 robot model at CoRL

NVIDIA released GR00T 2 at CoRL on November 9, a 2B-parameter Transformer based on Cosmos-Predict2.5-2B that combines camera footage, robot state, and text instructions to predict future views and action sequences, reporting a 40% zero-shot success rate on 50 tasks for which no training data existed in the target robots. Pretraining used 402,730 hours of first-person human video, 10,184 hours of handheld gripper demonstrations, and 23,500 hours of robot teleoperation data, and it released 100,000 hours of human data, Human-1M-100K, under CC BY 4.0. NVIDIA said the code will be released in December, and the whitepaper also states that the tasks evaluated for the 95% success rate with five demonstrations were filtered to those with at least an 80% success rate.

Google releases TPU 8tx and TPU 8i Max, expanding its eighth-generation TPU product line

Google is expanding its eighth-generation TPU product line with the addition of TPU 8tx for confidential inference and TPU 8i Max for ultra-long context and the highest per-rack compute density, forming a four-system lineup together with TPU 8t and TPU 8i. This is also the first time Google has launched four TPUs within a single year. TPU 8tx is based on the same chip as TPU 8t and is paired with a fifth-generation AMD EPYC CPU host, protecting data and model weights in use at the hardware level, with latency reduced by up to 2x compared with TPU 8t; TPU 8i Max packages two TPU 8i chips together, with compute, SRAM and HBM bandwidth all twice those of TPU 8i. Google also open-sourced TorchTPU, TPU Sync and the XLA Serving Engine, and announced the general availability of Cloud Pathways and GKE dynamic slicing.

• Google officially opens its Pathways distributed training system on Google CloudGoogle Developers Blog
• Google releases TPU 8i Max and TPU 8tx, doubling HBM bandwidthZhidongxi

Jensen Huang involved in an extramarital affair with an alleged Chinese spy, and Nvidia's board launches an investigation

Nvidia CEO Jensen Huang has admitted to an extramarital affair and is under investigation by the company's board of directors; The New York Times reported that U.S. intelligence officials believe Alice Huang, a Chinese businesswoman with whom he had an intimate relationship, is a Chinese spy. Nvidia's board said it currently has no reason to believe that U.S. national security or the company's confidential information has been compromised, and that Huang will remain in his position during the investigation. The revelation comes on the eve of a meeting between U.S. and Chinese leaders at the APEC summit; several lawmakers from both U.S. parties have called for a thorough investigation, while China's Foreign Ministry said the so-called Chinese spy narrative is purely groundless.

• Jensen Huang admits to an extramarital affair: when each party knew, and what will be investigatedGary Marcus：The Road to AI We Can Trust
• Yale professor Sonnenfeld says Jensen Huang's odds of staying or leaving are about 50-50, and the NVIDIA board investigation is rippling through the AI industrycnBeta
• Nvidia launches an independent investigation after Jensen Huang is alleged to be a target of foreign intelligenceThe Guardian Technology
• Nvidia's board launches an independent investigation into Jensen Huang's ties to a former Megaspeed executiveData Center Knowledge

Anthropic uses 1,000 Claude Fable 5.2 agents to reproduce a formal proof of Fermat's Last Theorem

Anthropic announced an experiment: 1,000 agents running an early version of Claude Fable 5.2 reproduced a Lean formal proof of Fermat's Last Theorem in 28.7 hours using only general-purpose tools and a shared Postgres database, with output tokens one-third of the original approach and a proof about one-seventh the length of the original. Each agent ran in a network-free sandbox, with shell commands and file edits first checked by a classifier, and separate offline monitoring reviewed team behavior. The team built its own tables in the database to divide the work, and for the hardest part, Langlands–Tunnell, it built a task tree of 2,418 tasks with a maximum depth of 10 layers.

ARC Prize publishes evaluation scores for Claude Fable 5.2 on the various ARC-AGI versions

ARC Prize published the scores of Anthropic's Claude Fable 5.2 on ARC-AGI: 81.2% under the ARC-AGI-3 standard test framework at a cost of $21.9K, 93.8% on ARC-AGI-2 at $0.98 per task, and 98.5% on ARC-AGI-1 at $0.36 per task. The 93.8% on ARC-AGI-2 is the third-highest model score it has seen, behind only GPT-6 Astra's 95.0% and GPT-6.1 Sol's 94.2%. At the max reasoning level, 28 of 55 semi-private games hit the 12-hour runtime limit, scoring 65.7%, below the 81.2% at the xhigh level.

• ARC Prize publishes evaluation results for Claude Fable 5.2 on ARC-AGIARC Prize