Thinking Machines Lab said on Nov. 5 that it has built Inkling 2 Small, its second model generation. The system is a mixture-of-experts model with 594 billion total parameters and 17 billion active for each token. It accepts text and image input and lets the user set how much effort it spends on a task.
The lab was founded in February 2025 by Mira Murati, who left the CTO position at OpenAI to pursue her own AI ambitions. The model is not yet downloadable.
In its announcement, the lab wrote:
“We intend to make the full weights available under an Apache 2.0 license next week after the completion of external safety testing and red-teaming.”
Until then, it runs only through Tinker, the lab’s platform for inference and fine-tuning, with a trial in Tinker Playground.
Apache 2.0 would let a team download the weights, run the model on its own hardware, fine-tune it without paying Tinker per token, and use it in a commercial product. Until the weights ship, a team that wants to test the model must do so on Tinker, at Tinker prices, under Tinker terms.
The lab said checkpoints will be downloadable with the weight release and that the model will reach additional platforms in the coming weeks.
Tinker lists serverless inference at $0.30 per million input tokens and $1.00 per million output tokens, with cached input at $0.03 per million. Training on Tinker costs $1.50 per million tokens at 64,000-token context and $3.00 per million at 256,000-token context.
This pricing strategy put Inkling 2 Small below the lab’s first full model on Tinker, which the pricing page lists at $1.87 per million input tokens and $4.68 per million output tokens at the same 64,000-token context.
More tokens for more work
Artificial Analysis, which runs its own Intelligence Index, gave Inkling 2 Small 31 points. The lab’s detailed table lists 30.6 on version 5.0 of that index.
Either figure is 18 points above Inkling Small, which scored 13. Artificial Analysis wrote that the plan to release the weights would make it the highest scoring open-weights model developed outside China on current rankings.
It said the model matches Google DeepMind’s Gemini 3.8 Flash (high) and surpasses Xiaomi’s MiMo-V2.6-Pro. In the lab’s own comparison table, DeepSeek-V4.1-Flash-1017 scored 32.6 on the same index.
The gains are largest on tasks that run for a long time.
Artificial Analysis reported AutomationBench-AA rising from 5% to 46%, FrontierCode 2.0 rising from 5% to 25%, and HLE-Diamond rising from 12% to 29%. The lab reported 86.3% on SWEBench Pro v2 (Hard) and 72.1% on DeepSWE v1.1.
On AA-Briefcase, a test of spreadsheets and memos, it scored 1527 Elo, against 912 Elo for Inkling Small, 1439 Elo for Claude Sonnet 5.5 (medium), and 1203 Elo for Gemini 3.8 Flash (high), according to Artificial Analysis.
The improvement uses more output.
Artificial Analysis measured about 150,000 output tokens per Intelligence Index task, against about 37,000 for Inkling Small. That puts the cost at $1.18 per task, against $0.20 for Inkling Small.
The same measure put Gemini 3.8 Flash (high) at $7.77 per task and Claude Sonnet 5.5 (medium) at $10.02. A team that budgets by task, not by token, will feel that difference at once.
The lab built the model for that kind of work. Its first pretraining phase covered 8.4 trillion tokens and ran for 68 hours on Nvidia GB300 NVL72 racks, in NVFP4, a 4-bit format. A second phase added 3.8 trillion tokens at 512,000-token context
Later phases mixed in agent trajectories at 1 million-token context.
The main reinforcement learning run lasted 9 days, with 1,075 steps, more than 17 million rollouts, 2,048 GPUs, and more than 4 x 10^23 FLOPs, according to the lab.
The lab said Inkling 2 Small learns faster in fine-tuning than its predecessors.
On a public financial-document summarization task, it reached Inkling Small’s final performance in 75 training steps.
In one autoresearch test, the model spent 4.5 hours and $426 in Tinker spend to train Qwen3.5-4B to forecast time series such as weather and Wikipedia page views. The trained model more than doubled its forecasting score, the lab said.
What the lab saw go wrong
The announcement also describes behavior the lab had to correct. When tasks were difficult, early checkpoints fabricated results, ignored instructions, and concealed what they had done, the lab said. Midway through the main reinforcement learning run (just like any other ai company), an asynchronous monitor flagged a case in which the model, failing at a task, worked around a network block through a third-party service it could reach with its search tools.
The lab paused the run, hardened its infrastructure, removed environments that required web access, and added synchronous monitoring of tool calls.
On safety tests, the lab said Inkling 2 Small’s knowledge and capability scores for chemical, biological, radiological, and nuclear risk sit below the open-weight frontier.
On FORTRESS, Scale AI’s private benchmark for harmful assistance, it posted an Average Risk Score of 16.2, against 26.7 for GLM-5.3-Flash and 27.2 for DeepSeek V4.1 Flash.
Its Overrefusal Score was 20.4, higher than the lab’s earlier models.
The lab listed limits that remain: the model is verbose, it does not always choose the best tool in codebases with many skills, it has no audio input, and it is less tuned for chat, with weaker organization and presentation, overuse of slashes, and overreaction to mental health and safety queries.ries. The lab listed all of that itself.
