Vercel, Sakana AI, Genspark, and Morph sign Microsoft open-weights letter

• The open-weights letter turned into a logo land grab: Vercel, Genspark, Morph, and Sakana AI all announced or showed support, with Vercel's post and Sakana AI's post adding fresh weight to the campaign.
• The policy hinge is distillation: the letter says policymakers should separate legitimate model-development techniques from misappropriation, a carveout highlighted by one excerpt and Natolambert's quote.
• Anthropic became the obvious target because it remained absent while OpenAI and Google appeared on the signatory page, and GergelyOrosz narrowed the major missing AI and cloud names to Anthropic and Amazon.
• The signup process looked improvised: one delivery-failure screenshot showed an EleutherAI-related attempt bouncing because the Microsoft group only accepted allowed senders.
• The safety argument stayed live: LechMazur's critique argued that open weights are irreversible once released, while Ryan Greenblatt's thread separated broad restrictions on released models from China-focused controls.

The current Microsoft letter lists Vercel, Sakana AI, Genspark, Morph, Google, OpenAI, Meta, Microsoft, NVIDIA, Mistral, Hugging Face, Ollama, LangChain, SpaceX, and dozens more. Its contact page still describes an earlier “over 20 companies” version and gives a form for new signatories. In the background, Moonshot’s Kimi K3 put the issue on a timer: DeepLearning.AI described it as a 2.8T-parameter model with weights promised by July 27, while CNN reported that Moonshot denied U.S. and Anthropic distillation allegations.

Vercel, Genspark, Morph, and Sakana AI were the cleanest new additions because each tied the letter to a concrete product or operating thesis.

• Vercel said open source, data, protocols, and research “enable the technological wonders we enjoy every day,” then called open weights the next frontier in its co-sign post.
• Genspark said open weights expand access, competition, and model choice for the “agentic future” it is building with Genspark 6.0 in its announcement.
• Morph said open weights matter because fair competition limits concentration, then credited Microsoft and Jensen Huang in its sign-on post.
• Sakana AI said open ecosystems support a healthy AI industry and technological sovereignty in its Japanese-language announcement.

Morph also made the policy stance operational: its Kimi K3 post said Kimi K3 would come to Morph via API and OpenRouter the next morning.

The sharpest sentence in the letter is the one that tries to rescue distillation from the “stealing” bucket.

The Microsoft letter defines distillation as using one model’s outputs to train or improve another, calls it “widely used” for improvement, evaluation, and validation, and says unlawful extraction from closed models should be handled through targeted legal and commercial frameworks.

That wording sits directly against the Kimi K3 fight. Kimmonismus's Axios excerpt said the Trump administration was backing open-weight models while accusing Moonshot of copying Anthropic’s Fable model through distillation, with officials distinguishing “legitimate AI distillation” from “large-scale, covert industrial distillation.”

Jensen Huang pushed the same broader framing in an Axios interview summarized by Rohan Paul's clip, saying AI systems will increasingly learn from other AI systems as the internet fills with AI-generated content.

Anthropic’s absence turned the letter from a policy statement into a social pressure campaign.

Julian Schrittwieser, posting as Mononofu, mocked NVIDIA and Microsoft’s open-source posture by asking for CUDA, GPU drivers, Windows, and Office to be opened in his screenshots. AndrewYNg answered that keeping one’s own code private is different from trying to stop others from opening theirs in his reply.

The broader criticism was less subtle. francoisfleuret accused Anthropic and Dario Amodei of lobbying against open-source AI, while QuixiAI's cancellation post showed a Claude Max cancellation email and said the reason was Anthropic’s stance.

There was a counterpoint inside the pile-on: Arohan's post reminded people that Anthropic had donated to open-source security, linking to the Python Software Foundation’s announcement of a $1.5 million, two-year contribution.

The letter spread faster than its intake process.

BlancheMinerva said EleutherAI was not invited, had no obvious way to sign, and had not heard back after DMs in one thread. A later screenshot from the same thread showed an email to openweights@microsoft.com bouncing because the group only accepted allowed senders.

The official Microsoft contact page now asks interested companies to submit name, company, email, and message. That page still lists the original smaller set of signatories, while the main letter page lists the expanded coalition.

The disagreement was not only open versus closed. It was also whether policy should treat released weights, Chinese model progress, distillation, and operational deployment as separate problems.

LechMazur argued that open weights borrow credibility from open-source software even when training data and code remain closed, and warned that released weights cannot be recalled in his critique. BlackHC made a narrower technical point in one reply: for open-weight models, safety has to be baked in because adversarial users can remove or bypass external controls.

Ryan Greenblatt split the policy target differently. His thread called broad restrictions on already released open-weight AIs bad policy, while supporting export controls and defenses against distillation aimed at slowing Chinese AI progress; his follow-up said government employees and defense contractors might reasonably be blocked from using adversary-trained open or closed models for automation.

The ecosystem argument had a live model menu behind it, not just slogans.

Sebastian Raschka’s roundup added six recent open-weight releases to his LLM Architecture Gallery via the linked gallery:

1. Nanbeige 4.2 3B: looped depth sharing, running the same 22-layer stack twice for 44 effective layers without duplicating weights.
2. Laguna S 2.1: 118B sparse MoE with 8B active parameters, 36 sliding-window layers, 12 global GQA layers, and a 1M-token context window.
3. Motif-3-Beta: 314B-A13B sparse MoE using Grouped Differential Latent Attention, inspired by Multi-head Latent Attention.
4. Solar Open 2: 250B-A15B hybrid MoE that interleaves three Kimi Delta Attention layers with one GQA layer.
5. Antares 1B: a small Cisco model built from the IBM Granite 4.0 1B backbone, post-trained with SFT plus GRPO for terminal cybersecurity tasks.
6. BTL-3: a rank-32 LoRA adapter for Qwen3.6-27B aimed at coding agents and structured tool use.