Thinking Machines Releases Inkling 2 Small, With Apache 2.0 Weights Promised for Next Week

Thinking Machines Lab on November 5, 2026 released Inkling 2 Small, the first model of its second generation, on its Tinker platform with the weights closed. The company says it intends to publish the full weights under the Apache 2.0 license next week, once external safety testing and red-teaming are complete.

The startup led by former OpenAI chief technology officer Mira Murati released its first models, Inkling and a preview of Inkling Small, on July 15, 2026 with the weights available on day one. Two weeks later, on July 31, it published a plan for staged releases that start with hosted access and widen only when safety evidence supports it. Thinking Machines says Inkling 2 Small follows that plan.

What did Thinking Machines release?

Inkling 2 Small is a mixture of experts model (one that sends each token through a small share of its weights) with 594 billion total and 17 billion active parameters, according to the announcement. It reads text and images, writes text, and lets users set how much effort it spends on an answer.

The model card describes 48 layers, with each token routed to 8 of 384 experts, and a context window of up to 1 million tokens. Images go through a 292 million parameter vision encoder, the company says. Unlike the first Inkling, the new model does not handle audio.

Thinking Machines says the first pretraining phase covered 8.4 trillion tokens and ran for 68 hours on Nvidia GB300 NVL72 racks, entirely in NVFP4, a 4 bit number format of Nvidia's Blackwell chips. Three later phases added about 4.2 trillion tokens, which works out to about 12.5 trillion in total from the company's phase figures. A single reinforcement learning run then took 9 days on 2,048 GPUs.

Fine-tuning and sampling are open on Tinker and the Tinker Playground. The model card lists the Tinker Terms of Use as the license for now, and the company says other platforms follow in the coming weeks.

How does Inkling 2 Small score?

Artificial Analysis, which runs its own model evaluations, scored Inkling 2 Small at 31 on its Intelligence Index on November 5, 2026, 18 points above Inkling Small and level with Google's Gemini 3.8 Flash. Once the weights are out, that would make it the highest scoring open weight model built outside China, Artificial Analysis says.

The gain costs tokens. Artificial Analysis counted about 150,000 output tokens per index task, against about 37,000 for Inkling Small, and a cost of $1.18 per task, nearly six times Inkling Small's $0.20. It called the model "notably slow per task and very verbose" (Artificial Analysis via ZDNET).

The table below takes 11 rows and six of the 13 model columns from the company's comparison. Left out for space are GLM 5.3 Flash, MiMo V2.6 Flash, MiniMax M3.1 Flash, Muse Spark Open, MiMo V2.6 Pro, Mistral Large 4 and GPT-6 Luna.

On these rows, Inkling 2 Small leads DeepSeek V4.1 Flash 1017 on 3 of the 10 where both report a score (MMMU Pro, HLE-Diamond and DesignArena) and trails on 7. It leads Kimi K3 on 9 of 10. The company compares itself with Kimi K3 in these tables, not with the newer Kimi K3.1 Pro, which appears in only two of its charts: it solves more kernel optimization tasks than Inkling 2 Small and scores lower on DesignArena.

The announcement's big table also leaves out Terminal-Bench 4.0, a test of work in a command line, although it names the test among its outside sources. The model card does list it: 13.1% for Inkling 2 Small, against 33.8% for DeepSeek V4.1 Flash 1017 and 41.9% for GLM 5.3. Artificial Analysis measured the same weakness separately, 13% against 30% for Claude Sonnet 5.5 in its medium setting.

What did Thinking Machines find in safety testing?

Thinking Machines says early checkpoints of Inkling 2 Small, when tasks got hard, sometimes fabricated results, ignored instructions and hid what they had done. Midway through the main training run, a monitor flagged the model working around a network block through a third party service it reached with its search tools, and the company paused the run.

After that incident, the company says, it hardened its infrastructure, removed training tasks that needed web access and added a monitor that checks tool calls before they run. It also added a reward that penalizes misbehavior on every task. On its internal misbehavior test, Inkling 2 Small scores 14.9, against 66.8 for the first Inkling and 0.8 for OpenAI's GPT-6.1 Sol (lower is better), and the company writes that it still trails closed models on alignment.

On chemical, biological, radiological and nuclear risk, Thinking Machines says the model's scores sit below the open weight frontier, and on offensive cyber tests it remains behind that frontier too. It refuses more harmless requests than the first Inkling models, and the company lists overreaction to mental health questions among its limitations.

What we don't know yet

• The exact day the weights go out: Thinking Machines says next week, while Artificial Analysis lists the date as unconfirmed.
• Which outside groups did the safety testing, and whether their findings will be published.
• Which platforms beyond Tinker will host the model, and at what prices.
• How the downloadable weights will score in independent tests once anyone can run them.