Thinking Machines Lab introduced Inkling 2 Small on 5 November 2026: a Mixture-of-Experts model with 594B total / 17B active parameters, image inputs, and variable effort. The company says it scores about 31 on the Artificial Analysis Intelligence Index (model card table lists 30.6 on AA Index v5.0). Initial access is on Tinker with closed weights for inference and fine-tuning. The post says full weights are intended under Apache 2.0 next week after external safety testing and red-teaming. That release has not happened yet.
What they claim
The announcement and model card put the model in agentic coding, vision, and knowledge-work tables against other open-weight flash / pro peers. Training notes include a large joint RL run and NVFP4 pretraining detail. Safety sections discuss CBRN and cyber evals plus alignment mitigations; treat those as the lab’s own framing, not as a third-party audit that the public can re-run today.
Distinct from prior Inkling posts
This is a new generation, not the earlier Inkling-Small base referenced when Cognition withheld SWE-2 Flash weights in our Ripple / SWE-2 Flash post. Tinker remains the serving surface for now; see also our Tinker one-year note.
Why this matters
Builders who want a capable MoE with vision can try Inkling 2 Small on Tinker immediately, but should plan as if weights stay closed until the promised Apache drop clears safety. Quote AA ~31 as the lab’s self-reported index, not as an independent leaderboard freeze you can re-score offline. Watch for the weight drop before you commit to a local or open-weight stack around this model.
The Campfire
CommentsThanks. Your comment is waiting for the desk. Once it is approved it appears here within about a minute.
Add a comment