Inkling 2 Small propels Thinking Machines to the top of non-China models

Thinking Machines Lab unveils Inkling 2 Small, a multimodal reasoning model leading non-China models according to Artificial Analysis. Its launch remains initially closed, with weight publication under Apache 2.0 announced later.

Thinking Machines Lab, the company founded by Mira Murati, has launched Inkling 2 Small, a multimodal reasoning model that takes first place among models built outside China in the Artificial Analysis Intelligence Index v5.0.1. With a score of 31, it surpasses several market references, but access begins in a proprietary way via Tinker and not as open weights. The promise of a later release of weights under the Apache 2.0 license changes the game for teams wanting to adapt a model without starting from scratch.

Un bond de performance qui ne passe pas inaperçu

The leap compared to Inkling Small is clear: the score increases by 18 points and the model produces about four times more output tokens per task. On the index, it matches Google DeepMind Gemini 3.8 Flash (high) and is ahead of Xiaomi MiMo-V2.6-Pro. On long and agentic tasks, where a model must chain steps without losing track, gains are even more visible. AutomationBench-AA rises from 5% to 46%, FrontierCode 2.0 from 5% to 25%, and HLE-Diamond from 12% to 29%.

The AA-Briefcase test, which measures office workflows with spreadsheets, presentations, and memos, illustrates the category change well. Inkling 2 Small reaches 1,527 Elo, compared to 912 for Inkling Small, and surpasses Claude Sonnet 5.5 (medium) at 1,439 Elo as well as Gemini 3.8 Flash (high) at 1,203 Elo. In this kind of exercise, the model no longer just responds: it keeps pace over multiple steps. This sometimes avoids too many meetings in IT services.

Coding, créativité et physique : le tableau est plus nuancé

Results are not uniform. In coding, Inkling 2 Small scores 25% on FrontierCode 2.0, better than Claude Sonnet 5.5 (medium) at 11%, but falls to 13% on Terminal-Bench 4.0, far behind Anthropic’s model at 30%. The message is clear: programming approaches differ, and benchmarks reward different reasoning styles.

On agentic creative tasks, the model surpasses Claude Haiku 5.5 and Grok 4.7 on AA-3D, and performs slightly better than Gemini 3.8 Flash on AA-3D and Code2Video-AA, though it remains behind Claude Sonnet 5.5 on these two tests. Compared to Inkling (xhigh), the leap is spectacular with 438 Elo more on AA-3D and 553 Elo more on Code2Video-AA. In sciences, the score also rises: 56% on CritPt v2, versus 34% for Inkling Small, though Gemini 3.8 Flash (high) remains ahead at 65%.

Moins d’hallucinations, mais surtout moins de tentatives

The most delicate point concerns hallucinations. On AA-Omniscience v1.1, the rate drops from 61% to 27%, which seems spectacular. However, Artificial Analysis specifies this decrease mainly comes from more cautious behavior: the attempt rate falls from 77% to 53%. Accuracy improves only slightly, from 32% to 34%, and the net score remains at 16, behind Gemini 3.8 Flash (high) at 23 and Claude Sonnet 5.5 (medium) at 19. In other words, the model answers less often when unsure. It’s more cautious, but not magic.

Un modèle lourd, cher et encore sous clé

Inkling 2 Small is not lightweight to run. The model has 594 billion parameters, including 17 billion active, and processes both text and images to produce text. On average, it uses about 150,000 output tokens per task on the Intelligence Index, versus 37,000 for Inkling Small, 31,000 for Claude Sonnet 5.5 (medium), and 99,000 for Gemini 3.8 Flash (high). This verbose appetite shows on the bill: $1.18 per task, nearly six times Inkling Small’s cost, but still less than Gemini 3.8 Flash (high) at $7.77 and Claude Sonnet 5.5 (medium) at $10.02 for comparable scores.

Pricing announced is $0.30 per million input tokens, $1 per million output tokens, and $0.03 per million cached input tokens. For now, the model remains accessible as a proprietary product via Tinker and Tinker Playground, before a full release of weights under the Apache 2.0 license announced later. Thinking Machines Lab presents this path as a way to let companies adjust and deploy their own models, rather than permanently renting access to a black box loaded with GPUs.

Key points

  • 31 points on the Artificial Analysis Intelligence Index v5.0.1.
  • 594 billion parameters, including 17 billion active.
  • AutomatioBench-AA rises from 5% to 46%.
  • AA-Briefcase reaches 1,527 Elo, ahead of Gemini 3.8 Flash.
  • Apache 2.0 license announced later by Thinking Machines.
  • Origin outside China: top of the ranking according to Artificial Analysis.

By the numbers

  • 31 — Inkling 2 Small’s score on the Artificial Analysis Intelligence Index v5.0.1.
  • 594 billion — model parameters, including 17 billion active.
  • $1.18 — average cost per task on the Intelligence Index.
  • $0.30 / million — input token price announced by Thinking Machines.
  • 150,000 tokens — average output per task for Inkling 2 Small.

To read