Thinking Machines Releases Inkling 2 Small: The Best Open Model Not Made in China

For the past couple of years, open models have had one main address: China. DeepSeek and Moonshot regularly topped the rankings and caught up with closed models from OpenAI and Anthropic at a fraction of the price. Now the West is getting an answer, and from two sides at once: Mistral's recent Le Chonk and the fresh Inkling 2 Small from Thinking Machines.

What the model is

Thinking Machines Lab is a small startup founded by Mira Murati, former chief technology officer of OpenAI. Inkling 2 Small has 594 billion parameters, accepts images, and responds only with text. The company focuses on web design, game design, and coding. It also stresses that during testing it evaluated not only the final code but also the model's own approach to the task. According to the developers, this yields a model that is stronger at code and works better inside agentic loops.

For now the model is formally proprietary. The weights are promised for next week. This scheme is becoming the norm: first several weeks of review by cybersecurity teams and reviewers, then publication. Mistral did the same with Le Chonk.

The numbers

Independent tester Artificial Analysis gave the model 31 points on its intelligence index. That is above average for its class and on par with Gemini 3.8 Flash. For scale: the leaders, Gemini 4 Carbon and Claude Opus 5.5, scored 60 and 59. So it is far from the top, but this is an open model, and by the rankings it will become the best open model from outside China.

Progress over the previous version is noticeable:

• automation grew by 41%;
• coding improved by 20%;
• the hallucination rate on AA-Omniscience v1.1 fell from 61% to 27%;
• accuracy rose from 32% to 34%;
• the share of attempted answers dropped from 77% to 53%.

The last three lines read like this: the model invents things less often because it more often honestly abstains from answering. Accuracy barely changed. For a product manager this is an important shift: "I don't know" is better than confident nonsense.

Limitations

The model has clear downsides. Artificial Analysis calls it noticeably slow per task and very verbose. This hurts both latency and the bill: verbosity means more output tokens. Input tokens cost $0.30 per million against a class median of $0.24, so slightly pricier. Also, no model wins everywhere, and it is too early to declare Inkling the best for coding.

What to do in practice

Until the weights are out, this is news to watch, not to migrate on. Once they appear, it is worth running the model on your own tasks: check the real cost given long answers, speed in agentic chains, and the refusal rate. For companies for which Chinese weights are inconvenient due to compliance or geopolitical requirements, this is the first serious Western option. For everyone else, the main takeaway is that competition among open models is intensifying, which means pressure on the prices of closed labs will grow too.