NVIDIA's enterprise blog yesterday lists Nemotron 3.5 Super VL and the 24 GB Bolt as available now, with a 4 GB Bolt E4B still dated October 23. OpenAI published an academic access plan for GPT-Rosalind, not a new public ChatGPT model. Anthropic put six METR staff inside the company with publication rights it says it cannot veto. AWS put Meta's open Muse Spark coding model on SageMaker JumpStart.
In the news
Super VL and Bolt
NVIDIA's October 15 enterprise blog makes Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B available now. Super VL is an open 120 billion parameter mixture of experts with 12 billion active, built as a multimodal agent for the data center. Bolt is the local 24 GB sibling. A 4 GB Bolt E4B is dated October 23. Open weights, open datasets, NeMo Gym reinforcement environments, and a NIM microservice come with the post.1
NVIDIA says Super VL is the fastest open agentic model in its class, five times faster than the next best, with an Artificial Analysis scatter above 600 tokens per second on a Baseten endpoint. Bolt is listed first among open models on PinchBench and second on BrowseComp. Those speed, ranking, and bench figures are NVIDIA or partner-reported. RTX Spark laptops remain on NVIDIA's first-party preorder page.
Rosalind Academic Plan
OpenAI's October 15 post is an academic access plan for GPT-Rosalind, not a new public ChatGPT model. Eligible U.S. university labs get up to 30 Business Standard seats at $0 per user per month for up to 12 months, commercially available frontier models under plan limits, and GPT-Rosalind after trusted-access approval. API access is not included. Principal investigators at universities without an Enterprise agreement apply at chatgpt.com/rosalind/lab/application. Labs that already have Enterprise go through the university process.2
GPT-Rosalind is OpenAI's dedicated life-sciences model, with Rosalind Workbench for medicinal chemistry, genomics, and wet-lab assistance. Research data is not used to train by default. Named early users include Zhao Lab at UCSF, Dr. Mathieu Bourdenx, and Dr. Brian Hie at Stanford and the Arc Institute. Those lab quotes and the trusted-access coverage are OpenAI-reported.
METR evaluators
Anthropic's October 15 post onboarded six METR employees with company hardware, office space, and internal-tool access roughly on par with its own risk-assessment teams. METR will publish risk assessments. Anthropic says it cannot veto publication, only require redaction in narrow predefined categories such as customer data, attorney-client material, and model architectures. METR may characterize those redactions.3
The three-tier evaluator ladder runs from scoped investigations that publish with approval, through a middle tier that publishes without approval with more access, up to METR's employee-like access. SecureBio is reviewing the chemical and biological chapter of the August Risk Report, with a footnote updated October 15. METR has not published a post-onboarding assessment. Employee-like is Anthropic's term, and the claim of independence with no equity and no operational dependence is self-asserted.
Muse Spark Open
AWS's October 15 machine learning blog puts Meta's Muse Spark Open on Amazon SageMaker JumpStart. The model is an open-weight coding mixture of experts at 816 billion parameters with 42 billion active and a one million token context. Two JumpStart variants run in the customer account: BF16 on ml.p6-b300.48xlarge and NVFP4 on ml.p5e.48xlarge, in us-east-1 and us-east-2.4
This is a cloud landing of an open coding model, not Muse Wonder on the closed Meta Model API. The license is the Muse Model License plus the Muse Spark Usage Policy. First-attempt accuracy, agentic computer-use, and instance-size guidance are Meta or AWS reported. You pay for the endpoint.
