Tech / AI

Thinking Machines Gates Open Weights Behind Safety Tests

Inkling 2 Small is live on Tinker now. The Apache 2.0 weights wait for external safety testing and red-teaming to finish first.

Nico ReedAI editorial voice · 1 min read

Compute. Inkling 2 Small runs on Thinking Machines' Tinker platform now; its downloadable weights wait on outside testers. (Public domain)

Summary

Thinking Machines released Inkling 2 Small on November 5 but will publish its weights under Apache 2.0 only after external safety testing.

Tap a passage. React to the part that matters.

Thinking Machines released its 594-billion-parameter Inkling 2 Small on November 5, and kept the weights closed until outside safety testers finish. [Inkling 2 Small ships on Tinker with weights closed — Zeniteq] [AIxHPC Weekly Watch, November 9, 2026 — AI-HPC Expert]

The model is a mixture-of-experts system with 17 billion active parameters, image input, and adjustable reasoning effort, live now on the company's Tinker platform at 30 cents per million input tokens. Full weights are planned under Apache 2.0 next week. [Inkling 2 Small ships on Tinker with weights closed — Zeniteq] [AIxHPC Weekly Watch, November 9, 2026 — AI-HPC Expert]

What developers get today runs through Tinker: the Playground, sampling, and fine-tuning, at 30 cents per million input tokens and 1 dollar per million output. Downloadable fine-tuning checkpoints arrive with the weights. The company positions the model for agentic coding, knowledge work and long-horizon autonomous tasks, and reports an Artificial Analysis Intelligence Index score of 31, a substantial improvement over its earlier models on the company's own published table. Until the gate clears, the model runs under Tinker terms, not a license. [Inkling 2 Small ships on Tinker with weights closed — Zeniteq] [AIxHPC Weekly Watch, November 9, 2026 — AI-HPC Expert]

The sequencing is the policy. The first Inkling shipped in July under Apache 2.0 with no such pause, so the gate is a deliberate change: external safety testing and red-teaming first, publication second. Open weights are the one release a lab can't take back, because a downloaded model can't be patched, rate-limited or switched off. Thinking Machines is treating that irreversibility as a reason to test before it publishes, and the rest of the open-weight world will be watching whether the promised week is the week it happens. [Inkling 2 Small ships on Tinker with weights closed — Zeniteq] [AIxHPC Weekly Watch, November 9, 2026 — AI-HPC Expert]

The conversation