Thinking Machines releases Inkling 2 Small, a 594B-parameter mixture-of-experts model, via its Tinker platform, with Apache 2.0 weights promised later

Thinking Machines has released Inkling 2 Small, a 594B-parameter mixture-of-experts model that uses 17B parameters at a time. It takes text and image input and has adjustable thinking depth. For now it is proprietary and available only on the company's Tinker platform. Apache 2.0 weights are promised once external safety testing finishes, with no date given, and more sizes are planned. The company reports an Artificial Analysis Intelligence Index score of 31 (up from 13) and large agentic gains, such as AutomationBench-AA rising from 5% to 46%. Cost is about $1.18 per task, nearly six times the predecessor, and the model uses about four times the output tokens.

Developers and enterprises get a much stronger agentic model from a Western lab, but only through a hosted service for now. Anyone who wants self-hosting or fine-tuning has to wait for weights that have not been released. The cost per task is much higher than the predecessor's, so teams should test on their own workloads before assuming the benchmark gains are worth the spend. If the Apache 2.0 licence does arrive, it would be a more permissive option than some rival open-weight licences, which matters mainly to firms with licensing constraints. Until then the practical impact is limited.

The "strongest Western open-weight model" claim describes a model that is not open-weight yet, and it rests on the company's own benchmark reporting. The "far lower cost per task than Gemini 3.8 Flash" comparison sits awkwardly beside a cost nearly six times higher than the predecessor's. The hallucination drop from 61% to 27% is less impressive when the response rate fell from 77% to 53%, since the model may simply be declining to answer more often. No primary links were supplied, so these figures are second-hand via The Decoder.

1. 01Whether the Apache 2.0 weights are actually published, with a date, and whether the licence text has no extra restrictions
2. 02Independent reproduction of the Artificial Analysis score of 31 and the agentic results, plus independent cost-per-task figures from third parties
3. 03Real-world pricing and uptake on Tinker, and whether the response-rate drop (77% to 53%) shows up as refusals or incomplete answers in developer reports