6 November 2026 · 2 min read
Thinking Machines releases Inkling 2 Small, a western open-weight model that challenges Chinese AI
The ex-OpenAI CTO's startup has launched a 594B-parameter model with strong coding and agent capabilities. Weights will be released next week, making it the highest-scoring open model built outside China.
Thinking Machines Lab, the startup founded by ex-OpenAI CTO Mira Murati, has released Inkling 2 Small, a new open-weight model that directly competes with leading Chinese open models [1]. The company plans to release the model's weights next week, following a staggered process that allows time for safety and security testing [1].
What Inkling 2 Small brings
The model packs 594 billion parameters and accepts image inputs, though it outputs only text [1]. Thinking Machines highlighted its web and game design strengths and said it evaluated the model's approach to coding problems alongside the code it generated during testing. "The result is a model that is both stronger at coding and easier to use inside real agent loops," the company stated [1].
On the Artificial Analysis Intelligence Index, a composite benchmark that tracks progress toward artificial general intelligence, Inkling 2 Small scored 31 – tying with Gemini 3.8 Flash and placing above average for its class [1]. For context, the top models on that index are Gemini 4 Carbon (60) and Claude Opus 5.5 (59) [1].
Performance and trade-offs
Artificial Analysis noted that Inkling 2 Small is "amongst the leading models in intelligence" and is reasonably priced compared to similar models, but it is "notably slow per task and very verbose" [1]. Its pricing sits at $0.30 per million input tokens, which is somewhat above the class median of $0.24 [1].
Compared to its predecessor, the model jumped 41% in automation capabilities and 20% in coding abilities [1]. Hallucination rates also improved significantly: on the AA-Omniscience v1.1 benchmark, the hallucination rate fell from 61% to 27%, while accuracy rose from 32% to 34% and the attempt rate dropped from 77% to 53% [1].
The open-weight landscape
If Thinking Machines follows through with releasing the weights, Inkling 2 Small would become "the highest-scoring open weights model developed outside China based on current rankings," according to Artificial Analysis [1]. This matters because Chinese labs like DeepSeek and Moonshot have dominated open-weight leaderboards, putting pressure on Western companies that rely on expensive closed models [1].
The release also follows French lab Mistral's Le Chonk model, another Western open-weight effort that now keeps pace with OpenAI's GPT-6.1 Terra in coding [1]. Both models use a staggered release: weights are held back for a few weeks while cybersecurity and safety reviewers test the model [1].
What it means for developers
For builders, Inkling 2 Small offers a new option that balances strong coding performance with open-weight access. The model's design focus on agent loops suggests it may be easier to integrate into automated workflows, though the reported slowness and verbosity could affect latency-sensitive applications. The higher-than-median token pricing may also be a consideration for cost-conscious projects.
The jump in coding and automation scores, combined with a much lower hallucination rate, indicates that the model is more reliable for tasks where factual accuracy matters. Once the weights are available, developers will be able to run and fine-tune the model on their own infrastructure, avoiding per-token costs entirely.
Sources
- open-weight
- ai model
- coding
- thinking machines
- inkling 2
Researched and drafted with AI from the sources above.
Learn to build with AI
Hands-on tutorials on Cursor, Claude Code, AI agents and more.