Thinking Machines Aims for the Top of Open Weight Models

Key Takeaways

1Inkling 2 Small from Thinking Machines scores 31 on the Artificial Analysis index, matching Gemini 3.8 Flash

2The model has 594 billion parameters, accepts images as input, and shows progress in automation, coding, and accuracy

3Thinking Machines plans to release the weights next week, a practice that is becoming increasingly common

4Artificial Analysis considers the model to be effective but slow, verbose, and more expensive than the median in its category

Why it matters : Inkling 2 Small could become the highest-rated open-weight model developed outside of China, marking a significant advancement for Western players against Chinese dominance.

With Inkling 2 Small, Thinking Machines aligns itself with current benchmarks while planning to release its weights next week. Third-party evaluations place it on par with Gemini 3.8 Flash on a composite index, and it is anticipated to become the highest-rated model developed outside of China once released, subject to current rankings.

Ranking: a score of 31 and an edge outside of China mentioned

Artificial Analysis credits Inkling 2 Small with a score of 31 on its composite index. This level is above the average for its category and equal to Gemini 3.8 Flash. The top spots on this index remain occupied by Gemini 4 Carbon and Claude Opus 5.5, with scores of 60 and 59, respectively. Artificial Analysis further estimates that the release of the weights would make Inkling 2 Small the highest-rated open-weight model developed outside of China according to current rankings.

Measured capabilities and progress since the previous model

Inkling 2 Small totals 594 billion parameters. It accepts images as input and generates only text. Thinking Machines highlights applications in web design and gaming, stating that it has evaluated the model's approach to coding problems as well as the quality of the produced code, claiming improvements in coding and easier integration into agent loops. Compared to its predecessor, Artificial Analysis notes gains of 41% in automation and 20% in coding. The organization also observes improvements in accuracy: on AA-Omniscience v1.1, the hallucination rate drops from 61% to 27%, accuracy rises from 32% to 34%, and the trial rate decreases from 77% to 53%.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Weight release announced and industry practices

Thinking Machines plans to release the weights of Inkling 2 Small next week. The staggered availability of weights is becoming more common, with testing phases by cybersecurity teams and security reviewers lasting several weeks before the release. Mistral recently followed this approach for Le Chonk, a model that now keeps pace with OpenAI's GPT-6.1 Terra in coding, a rare feat for an open model developed in the West. Le Chonk is also described as excellent in coding.

Chinese competition, costs, and reported limitations

Chinese players have dominated open models in recent years, with systems like those from DeepSeek and Moonshot regularly ranking at the top, competing with more expensive closed models from OpenAI and Anthropic. Artificial Analysis qualifies Inkling 2 Small as among the most capable in intelligence, while describing it as slow per task and very verbose. The pricing is considered somewhat high at $0.30 per million input tokens, compared to a median of $0.24 in its category. The comparison index aggregates scores in coding, sciences, and general domains, and the specialization by tasks makes it difficult to designate a single best overall coding model.