Meta Releases Muse Spark Open as Anthropic Embeds METR Evaluators and OpenAI Brings GPT-Rosalind to Academic Labs
Open weights, outside oversight, and science access arrived in the same news cycle. Meta released Muse Spark Open, an 816-billion-parameter multimodal model sized to run across two Mac Studios. Anthropic gave a METR team employee-like access, moved Code Review to general availability, and dropped the Opus 5.5 fast mode waitlist. OpenAI opened sponsored GPT-Rosalind seats to U.S. university labs and shipped GPT-Live-1.1, the Simons Foundation launched a $15 million AI-for-math effort with both labs, and NVIDIA stretched Nemotron 3.5 from the data center to 24GB machines.
Meta Releases Muse Spark Open, an 816B Open-Weight Model Sized for Two Mac Studios
Meta released the weights for Muse Spark Open on October 15. The Meta AI Research announcement introducing Muse Spark Open describes a natively multimodal model with 816 billion total parameters, 42 billion active, and a one-million-token context window, built for coding and long-horizon agentic work. Meta calls it the most capable US open-weight model to date and says it is the first in its class to run on a desktop, using two Apple M5 Ultra Mac Studios with MLX. Hosted APIs are live at Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI, with weights on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry Models.
"We believe the future is for everyone, and that includes the freedom to run advanced models wherever, including at home." — Meta AI Research
The Muse Spark Open model card and license on Hugging Face add the fine print. Self-hosting needs at least 700GB of GPU memory for the NVFP4 checkpoint and 2TB for BF16. The Muse Model License requires a “Built with Muse Spark Open” notice, and API resellers with more than $20 million in trailing twelve-month gross revenue must sign a separate agreement with Meta. In Meta’s own table, Muse Spark Open scores 39 on the Artificial Analysis Intelligence Index, behind GLM 5.3 at 45, Kimi K3 at 44, and Qwen3.8-2.4T-A95B at 40, so the “US” qualifier carries real weight. Meta says the model met its Superintelligence Scaling Framework thresholds, with a full evaluation report still to come.
Muse Spark Open turns on-premises frontier-class coding agents into a hardware budgeting question. Run the license past counsel first, since the attribution notice and the $20 million model-as-a-service threshold matter for anyone reselling inference. Then test it on your own repositories against the higher-scoring open models in Meta’s table, with human confirmation for irreversible actions.
Anthropic Gives METR Employee-Like Access as Its Top Evaluator Tier
Anthropic said on October 15 that it has started giving employee-like access to a team from METR, the nonprofit that evaluates frontier AI models. According to Anthropic’s update on independent evaluators, six METR employees now have company hardware, office space, and internal tool access roughly on par with Anthropic’s own risk assessment teams. METR will publish its analysis of Anthropic’s risk levels, incidents, or practices. Anthropic cannot veto publication but can require redactions in narrow predefined categories, such as customer data, privileged legal material, and model architecture, and METR may say publicly whether it thinks those redactions were justified.
"We aim to close the gap between what evaluators understand about a frontier AI company’s systems and what the company itself understands." — Anthropic
METR sits in the deepest of three planned tiers. First-tier evaluators publish only with Anthropic’s approval, and second-tier evaluators can publish without it. SecureBio is reviewing the chemical and biological weapons chapter of Anthropic’s August Risk Report, a broader partnership with Accenture and Faculty is in progress, and early talks include Hugging Face.
"We’re embedding evaluators voluntarily, and we believe every frontier AI company should do the same." — Anthropic
Risk and procurement teams now have a sharp due-diligence question for every frontier vendor: which outside evaluator sits at which tier, and can it publish without your approval? Anthropic’s no-veto rule and publicly contestable redactions set a usable bar. The first METR publication will show whether embedded evaluation produces evidence a board can rely on.
Claude Code Review Reaches General Availability as Opus 5.5 Fast Mode Drops Its Waitlist
Anthropic moved Code Review out of research preview for Claude Team and Enterprise organizations. The Claude announcement that Code Review is generally available says most reviews now finish about three times faster than in preview, roughly 6 to 10 minutes instead of about 20, and admins choose Quick, Standard, or Deep per repository. Eligible Enterprise organizations can claim at least $10,000 in credits by January 15, 2027, expiring 30 days after the claim. The Claude Code documentation for Code Review lists the limits: GitHub only, and no availability for organizations using Zero Data Retention, the HIPAA configuration, or customer-managed encryption keys. A review never approves or blocks a pull request.
"Coding agents keep increasing the pace at which teams write code, and the review bottleneck hasn't eased." — Anthropic
Anthropic’s developer account also said Opus 5.5 fast mode on the Claude Platform is out of research preview with the waitlist removed, with Google Cloud coming soon. The Claude Platform fast mode documentation promises up to 2.5 times higher output tokens per second from the same model, at $8 per million input tokens and $40 per million output tokens, on the Claude API and Azure-hosted Microsoft Foundry. The gain is in output speed, not time to first token. Fast mode is not offered with the Batch API, Priority Tier, or Claude Platform on AWS, and switching speeds invalidates the prompt cache.
Budget these separately. Pilot Code Review on one high-churn repository at Standard before tuning others, after confirming you are not a ZDR, HIPAA, or CMEK organization. Reserve fast mode for streaming turns where a person is waiting, and measure the cache-miss cost of switching speeds before routing per request.
OpenAI Offers Sponsored GPT-Rosalind Seats to U.S. University Labs
OpenAI introduced the Rosalind Academic Plan on October 15. Under OpenAI’s announcement of the Rosalind Academic Plan, eligible labs at U.S. universities can receive up to 30 Business Standard seats at $0 for up to 12 months, plus GPT-Rosalind, OpenAI’s life sciences model, after trusted-access approval. API access is not included, and research data shared through the plan is not used for training by default. Before activation, OpenAI verifies the institution and the lab lead, reviews intended use, and requires a designated institutional safety contact, with monitoring continuing afterward.
"Rosalind will not only help us improve our current discovery and design tasks, but also apply them at larger scales than previously possible while still providing useful insights to biological researchers." — Dr. Brian Hie, Stanford University and Arc Institute
Research offices should treat the Rosalind plan as a compliance onboarding, not a software giveaway. Identity checks, a named safety contact, and ongoing monitoring mean someone must own the relationship, and the lack of API access limits pipeline automation. Industry partners working with these labs should settle data-handling terms early.
Simons Foundation, Anthropic, and OpenAI Commit $15 Million to Help Mathematicians Navigate AI
The Simons Foundation announcement of its $15M effort for mathematicians in the AI era says the foundation, Anthropic, and OpenAI will commit $15 million in combined funding and resources. The foundation will manage and award the money, including through an open call for proposals, for workshops and initiatives that help researchers evaluate AI-assisted results and shape standards for AI use. Both labs will also provide tools and compute to Simons-supported researchers. OpenAI’s post on its $5 million Simons contribution and Anthropic’s post on its $5 million Simons commitment each put their share at $5 million.
"The community should be driving discussions of how these tools are evaluated, adopted and used, not any single company or institution." — David Spergel, President, Simons Foundation
The key design choice is governance: the foundation, not the labs, controls the money and the call for proposals. Institutions accepting lab funding for AI programs should copy that pattern. For math departments, the practical first step is funding verification workshops so faculty can check AI-generated proofs before they reshape publishing norms.
NVIDIA Expands Nemotron 3.5 From Data Center to Desktop
NVIDIA released Nemotron 3.5 Super VL on October 15, an open 120-billion-parameter model for always-on agents. The NVIDIA blog on Nemotron 3.5 Super VL and Bolt claims accuracy comparable with NVIDIA’s largest reasoning model at about one-fifth the parameters and five times the speed of the next best open model in its class, with vision tuned for OCR, tables, charts, web pages, and PDFs. Weights, datasets, and training recipes are open. Nemotron 3.5 Bolt 74B-A9B fits in 24GB for local agents and is available now, and the 4GB Bolt E4B arrives October 23. Separately, the NVIDIA and Mercedes-Benz press release on their expanded automated-driving partnership said on October 16 that the companies will jointly develop next-generation driver assistance and automated driving on NVIDIA Hyperion, Alpamayo, and Halos.
"Any organization that wants to run agents on NVIDIA infrastructure now has an open, optimized model for their tier, without having to choose between capability and deployability." — NVIDIA
Nemotron’s tiering offers one open family from server to laptop, which simplifies evaluation and fine-tuning pipelines. Benchmark Super VL on your own document-heavy work, such as invoices, claims, and contracts, where NVIDIA aims its vision claims. Treat the fivefold speed figure as vendor-reported until you reproduce it.
OpenAI Ships GPT-Live-1.1 for ChatGPT Voice, With Disclosed Safety Regressions
OpenAI began rolling out GPT-Live-1.1 and a smaller GPT-Live-1.1 Quip in ChatGPT Voice on October 15. The OpenAI Help Center ChatGPT release notes on GPT-Live-1.1 cite more natural accents and new voices for French, German, Hindi, Japanese, Korean, and Mexican Spanish. OpenAI’s announcement post for GPT-Live-1.1 says it also reaches Voice in Codex and dots, and is coming soon to the API. The GPT-Live-1.1 system card from OpenAI reports parity with the predecessors on production-prompt safety evaluations, but regressions on harder synthetic prompts.
"Analysis shows that GPT-Live-1.1 has statistically significant regressions on sexual, mental health, emotional reliance, and gore categories compared to GPT-Live-1; more research is needed to understand the regressions, and we plan to improve in these categories in future releases." — OpenAI, GPT-Live-1.1 System Card
Teams deploying ChatGPT Voice in wellness, education, or support settings should read the system card before the release notes. Mental health and emotional reliance are where voice carries the most weight, and OpenAI says these evaluations exclude protections like Trusted Contact and Parental Controls. Confirm those controls are on, and rerun your own spoken red-team prompts.
Today’s thread is access with conditions. Meta and NVIDIA opened weights spanning an 816B model down to a 24GB local model, with license and verification caveats attached. Anthropic let an outside evaluator work almost like staff and publish without a veto, and turned two preview features into products with explicit eligibility limits. OpenAI paired sponsored science seats with identity-verified gates and published a voice system card naming its own regressions, while the Simons partnership leaves mathematicians in charge of judging AI results. Read licenses, system cards, and evaluator terms as closely as benchmark tables.
Need help navigating AI for your business?
Our team turns these developments into actionable strategy.
Contact SEN-X →