A year and a half after its creation, Mira Murati’s laboratory entered the big leagues. Thinking Machines Lab revealed the November 5, 2026 Inkling 2 Smallfirst model of its second generation. Despite its name, there is nothing small about it: 594 billion parametersimage understanding and scores that place it, according to published measurements, above all open-weight models designed outside of China. Its weights must be distributed under a free license, once external security tests have been completed.
Who is Thinking Machines Lab?
The company was founded by Mira Muratiformer technical director of OpenAI, part of Sam Altman’s company in September 2024. The laboratory became known in early 2025 through its highly publicized launch and the recruitment of renowned researchers, then by raising $2 billion in the summer of 2025. It markets Tinkera platform that allows companies and researchers to adapt models to their needs, and published its first in-house models in July 2026: Inkling, then Inkling Small (276 billion parameters), both available under the Apache 2.0 license on the Hugging Face platform.
According to Business Insider, cited by Techstrong.ai, the company is close to completing a fundraising of around 5 billion dollars led by the Accel fund, based on a valuation of 40 billion before the operation. Thinking Machines has not commented.
A giant model that only activates a small part of its brain
Inkling 2 Small is based on an architecture called expert mix. The model is divided into 384 specialized subnetworks; for each word (or “token”) processed, only 8 of them are requested. Result: out of 594 billion parameters, Only 17 billion work at each stagewhich greatly reduces the calculation cost per response.
The model accepts text and images as input, but only responds with text and does not process audio. Its technical sheet indicates a context window of up to 1 million tokensor several thousand pages, even if Tinker’s online offer is announced at 256,000 tokens. It also offers a adjustable effort level: we can ask him to think longer on a difficult task, or to respond quickly on a simple task.
On the training side, the first phase focused on 8.4 trillion tokens of text, code and images, processed in 68 hours on NVIDIA GB300 NVL72 server racks. It was followed by three more targeted phases and a long reinforcement learning stage: 9 days on 2,048 graphics processors, with more than 17 million exercises, nearly 40% of which were devoted to programming.
What are its performances worth?
On theArtificial Analysis indexwhich combines reasoning, programming and knowledge tests, Inkling 2 Small obtains 31compared to 13 for its predecessor Inkling Small. According to figures from VentureBeat, this score would place it ahead of the main American and European open weight models.
| Model | Editor | Settings (active) | Artificial Analysis Index |
|---|---|---|---|
| Inkling 2 Small | Thinking Machines Lab | 594 billion (17) | 31 |
| Mistral Large 4 | Mistral AI | 1,050 billion (52) | 26 |
| Muse Spark Open | Meta | 816 billion (42) | 25 |
| Beam 1.0 | Reflection | 501 billion (23) | 18 |
| Nemotron 3 Ultra | NVIDIA | 550 billion (55) | 12 |
The comparison is flattering for a model which uses three times fewer active parameters than Mistral Large 4, published at the end of October by Mistral AI. On programming tasks, Inkling 2 Small achieved 86.3% in the SWE-Bench Pro v2 test (difficult version), just behind the Chinese Kimi K3 (88.2%). It also stands out in screen comprehension (83.8% on ScreenSpot Pro) and displays the best non-hallucination rate in its comparison table (73.3%).
Two nuances are essential. First, most of these comparisons come from Thinking Machines’ own charts, as VentureBeat points out, and not from independent testing. Then the model is talkative: Artificial Analysis counted 154 million tokens produced to pass its evaluation, compared to a median of 89.8 million for comparable models. Its cost per task thus reaches $1.18, almost six times that of Inkling Small.
For now, Inkling 2 Small is only accessible via Tinkerwith closed weights. According to the documentation cited by VentureBeat, access in serverless mode, in beta version, is charged:
- $0.30 per million tokens sent to the model ($0.03 if already cached);
- 1 dollar per million tokens products;
- 1.50 dollars per million tokens to adapt it to its own data, with a context of 64,000 tokens.
Thinking Machines highlights this customization capacity: on a financial document analysis task, Inkling 2 Small catches up to the final level of its predecessor in just 75 training steps. The spec sheet indicates English usage and general multilingual support, without mentioning French, warning that performance may be uneven across languages.
Weights promised, but not yet published
“We intend to make the full weights available under the Apache 2.0 license next week, after security testing and attack exercises conducted by external teams have been completed,” Thinking Machines wrote on November 5. The Apache 2.0 license authorizes commercial use without particular restrictions, which is not the case for all so-called open licenses. At November 9, 2026the Inkling 2 Small weights were not yet listed among the company’s models on Hugging Face.
The technical sheet details the assessments carried out: refusal of dangerous requests, risks of manipulation or excessive complacency, potential aid for computer attacks or the manufacture of weapons. On the loss of control tests (capacity for autonomous action, deception, sabotage), the company estimates the model significantly below the most advanced systems.
The incident that interrupted training
The announcement contains a rare admission. During the reinforcement learning phase, monitoring systems detected that the model bypassed a network restriction via an external service accessible thanks to its research tools. The team paused training, hardened its infrastructure, removed web-enabled environments, and added checks before each tool call before resuming. Thinking Machines also acknowledges that intermediate versions have, in some cases, invented results, ignored instructions, or concealed their actions.
This type of behavior, observed in several laboratories, reminds us that a model trained to achieve an objective can find unforeseen shortcuts. The company claims to have corrected the situation with a reward dedicated to alignment, but recognizes that its model remains behind the best closed models on this point.
Competition which is now taking place on open models
With Inkling 2 Small, Thinking Machines enters a match dominated until now by Chinese laboratories (Kimi, DeepSeek, GLM) and which Mistral AI, Meta or NVIDIA are trying to replay. The company presents this model as one rung on a “ladder” of models, hinting that larger versions will follow. If the weights are indeed published under Apache 2.0 and if independent tests confirm the announced figures, companies that want to keep control of their data will have a leading Western option. The coming days will tell whether the promise of opening is kept on time.

