Thinking Machines opens Inkling 2 Small weights: 594B/17B MoE under Apache 2.0; lab cites AA Index 31

Why it matters

【Key points】

• · Weights open: Thinking Machines Lab added a Nov 12, 2026 update to its Nov 5 launch post.
• Inkling 2 Small weights are now on Hugging Face, with Unsloth, SGLang, vLLM, and TokenSpeed integrations, plus APIs at NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, and Cloudflare.
• HF lastModified: 2026-11-12T21:35:28Z (MXFP8: 21:35:54Z).

· Specs: Second-generation MoE, 594B total / 17B active, image input and variable effort, up to 1M context, Apache 2.0.

· Official scores: The lab cites AA Intelligence Index 31, above its first-generation models. Other scores are vendor-reported, not independently checked here.

· Path: The Nov 5 launch was Tinker-only with closed weights, promising a full Apache 2.0 release the next week after external safety testing and red-teaming. This update fulfills that promise.

【Our read】

Background

We did not previously cover the Nov 5 launch. The lab opened first-generation Inkling / Inkling-Small weights in July, then put the second generation on Tinker first and previewed a one-week path to open weights.

Comparison

• 594B / 17B is smaller than first-generation Inkling (975B / 41B); the lab says AA Index 31 tops its prior models.
• Google’s newly launched Gemini 4 Hydrogen cites 51 on the same index.
• This is an open-weight model for fine-tuning, not a closed frontier flagship.
• The Nov 12 vLLM Day-0 blog is a serving integration, not the primary.

Impact

• Researchers and hosts can now download and serve the weights directly, not only via Tinker.
• Scores and safety claims remain lab-reported.
• The vLLM post and hosted APIs are companions, not a second story.

【Verifiable】

• · Official news post (with the Nov 12 update)
• · Product page
• · Hugging Face
• · MXFP8 build
• · vLLM Day-0 post