Inkling 2 Small aims to climb to the top of open models

Thinking Machines unveiled Inkling 2 Small, a proprietary model at launch, with weights promised later. Its score of 31 on the Artificial Analysis Intelligence Index places it among the best open models outside China [to be verified].

Thinking Machines has launched Inkling 2 Small, a reasoning model initially available via Tinker and Tinker Playground, with weights promised later. On the Artificial Analysis Intelligence Index v5.0.1, it scores 31 points, 18 more than Inkling Small. At this stage, the promise of release under the Apache 2.0 license would make it the highest-ranked open model outside China, according to Artificial Analysis [to be verified].

Des gains nets sur les tâches d’agent, pas seulement sur le papier

The performance leap is particularly noticeable on so-called agentic uses, that is, when the model chains several actions to accomplish a task. AutomationBench-AA rises from 5% to 46%, FrontierCode 2.0 from 5% to 25%, and HLE-Diamond from 12% to 29% compared to Inkling Small. On AA-Briefcase, which tests long desktop workflows with spreadsheets, presentations, and memos, Inkling 2 Small reaches 1527 Elo versus 912 for Inkling Small, 1439 for Claude Sonnet 5.5 (medium), and 1203 for Gemini 3.8 Flash (high).

Pourquoi le modèle parle autant

Inkling 2 Small also produces many more output tokens, those text units generated by the model. It outputs on average about 150,000 per task on the Intelligence Index, compared to about 37,000 for Inkling Small, 31,000 for Claude Sonnet 5.5 (medium), and 99,000 for Gemini 3.8 Flash (high). As a result, the cost rises: $1.18 per task, nearly six times the cost of Inkling Small, but still less than several comparable models in score. The model has not only learned to respond better; it also talks more. Like a very motivated colleague who hasn’t yet found the “summarize” button.

Un modèle plus prudent, donc moins halluciné

On AA-Omniscience v1.1, Inkling 2 Small significantly reduces its hallucination rate, dropping from 61% to 27%. The improvement partly comes from more selective behavior: the model attempts a response less often, with a try rate reduced from 77% to 53%, while accuracy modestly improves from 32% to 34%. Its net score rises to 16, compared to -9 for Inkling Small, but remains behind Gemini 3.8 Flash (high, 23) and Claude Sonnet 5.5 (medium, 19).

Le positionnement compte presque autant que le score

Thinking Machines launches Inkling 2 Small as a proprietary model, initially available on Tinker and Tinker Playground, then with weight publication at an unconfirmed date. The announced features are 594 billion total parameters, including 17 billion active, with text and image input, and text output. The listed prices are $0.30 per million input tokens, $1 per million output tokens, and $0.03 per million cached input tokens. In a market where rankings change quickly, the real question is now less “who wins today?” than “at what price, and for which uses.”

Key points

  • 31 points sur l’Artificial Analysis Intelligence Index v5.0.1.
  • Thinking Machines prévoit des poids Apache 2.0, date non confirmée.
  • AutomationBench-AA grimpe de 5 % à 46 %.
  • AA-Briefcase atteint 1527 Elo, devant Gemini 3.8 Flash.
  • AA-Omniscience v1.1 tombe à 27 % d’hallucination.

By the numbers

  • 18 points — gain d’Inkling 2 Small sur Inkling Small, Artificial Analysis, 5 novembre 2026.
  • 594 milliards — paramètres totaux annoncés pour le modèle, Artificial Analysis, 5 novembre 2026.
  • 17 milliards — paramètres actifs annoncés, Artificial Analysis, 5 novembre 2026.
  • 1,18 dollar — coût par tâche sur l’Intelligence Index, Artificial Analysis, 5 novembre 2026.
  • 150 000 tokens — moyenne de sortie par tâche, Artificial Analysis, 5 novembre 2026.

To read