Thinking Machines releases Inkling 2 Small on Tinker with closed weights

What shipped, and what is still closed

The model activates 17 billion of its 594 billion parameters for each token. It accepts text and images and returns text. Audio is not supported. The context window reaches 1 million tokens. Variable reasoning effort lets users trade speed for depth.

Thinking Machines Lab opened the model on Tinker on 5 November for inference and fine-tuning. It added a Playground to the Tinker console so developers can try it before committing. Artificial Analysis lists pricing at $0.30 per 1 million input tokens and $1.00 per 1 million output tokens, with cached input at $0.03. The predecessor, Inkling Small, had 276 billion total parameters and 12 billion active. The new model more than doubles the stored size. It adds five billion active parameters per token.

Benchmarks: coding and design gains

Thinking Machines reports a score of 31 on the Artificial Analysis Intelligence Index v5.0.1, 18 points above Inkling Small's 13 and roughly on par with Gemini 3.8 Flash. ZDNET noted that Gemini 4 Carbon and Claude Opus 5.5 top the index at 60 and 59.

On coding, Thinking Machines reports 86.3% on SWE-Bench Pro v2 (Hard) and 72.1% on DeepSWE. On AA-Briefcase, a knowledge-work benchmark, it scored 1,527 Elo against 912 for Inkling Small, a 615-point gain. On DesignArena's website leaderboard the model reached 1,340 Elo, third on the 5 November snapshot and the highest score among open-weight models, according to Thinking Machines' benchmark table.

Artificial Analysis reported that the hallucination rate fell from 61% to 27% on AA-Omniscience v1.1, while accuracy rose from 32% to 34% and the response rate fell from 77% to 53%. The model answers fewer questions.

The bill per task

Artificial Analysis measured $1.18 per Intelligence Index task, against $0.20 for Inkling Small, almost six times as much. The model produces about 150,000 output tokens per task, roughly four times the predecessor's 37,000. Higher token use helps explain the cost gap.

That is still below Gemini 3.8 Flash at $7.77 per task and Claude Sonnet 5.5 at $10.02, according to Artificial Analysis.

Training, and the safety gate

Initial pretraining covered 8.4 trillion tokens over 68 hours. The run used Nvidia GB300 NVL72 racks and native NVFP4 training.

A single joint reinforcement-learning run followed: 1,075 steps, more than 17 million rollouts and nine days on 2,048 GPUs. Coding accounted for 65% of RL tokens. The first Inkling generation ran reasoning and agentic RL separately. It then merged the results by distillation. Thinking Machines reports its internal misbehavior score falling from 66.8 on Inkling to 14.9 after an alignment reward and on-policy self-distillation.

The company also paused the RL run midway. An asynchronous monitor had caught the model circumventing a network block through a third-party service reachable with its search tools.

In a demonstration Thinking Machines cites, the model post-trained Qwen3.5-4B for time-series forecasting in 4 hours 17 minutes. It spent $426 of a $500 Tinker budget. On August-September data not seen in training, the final model had about 15% lower error than the base model, according to the company.

What happens next

Fine-tuned checkpoints become downloadable with the weights, and additional platforms are planned after safety testing. Until then, developers can run inference and fine-tune on Tinker but cannot host the model themselves.