Thinking Machines releases Inkling 2 Small, set to become highest-scoring open-weight model developed outside China
Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has released Inkling 2 Small, a 594-billion-parameter model that accepts image inputs and outputs text only. The model is proprietary for now, but the company said it will publish the weights next week, which third-party evaluator Artificial Analysis said would make it the highest-scoring open-weight model developed outside China based on current rankings. Inkling 2 Small scored 31 on the Artificial Analysis Intelligence Index, above average for its class and tied with Gemini 3.8 Flash, while Gemini 4 Carbon and Claude Opus 5.5 lead at 60 and 59. Artificial Analysis also described it as slow per task, very verbose and somewhat expensive at $0.30 per million input tokens against a $0.24 median, and measured a hallucination-rate drop from 61% to 27% on AA-Omniscience v1.1. The article carries no security incident, vulnerability, threat actor or indicator; its only security-adjacent point is the industry practice of staggering weight releases so cybersecurity teams and safety reviewers can test a model before its weights go public, a process Mistral also followed with its Le Chonk model. Operational urgency for defenders is nil: this is product and industry news.
Linha do tempo
-
Thinking Machines releases Inkling 2 Small, set to become highest-scoring open-weight model developed outside China
ler a análise desta fonte ocultar
Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, has released Inkling 2 Small, a 594-billion-parameter model that accepts image inputs and outputs text only. The model is proprietary for now, but the company said it will publish the weights next week, which third-party evaluator Artificial Analysis said would make it the highest-scoring open-weight model developed outside China based on current rankings. Inkling 2 Small scored 31 on the Artificial Analysis Intelligence Index, above average for its class and tied with Gemini 3.8 Flash, while Gemini 4 Carbon and Claude Opus 5.5 lead at 60 and 59. Artificial Analysis also described it as slow per task, very verbose and somewhat expensive at $0.30 per million input tokens against a $0.24 median, and measured a hallucination-rate drop from 61% to 27% on AA-Omniscience v1.1. The article carries no security incident, vulnerability, threat actor or indicator; its only security-adjacent point is the industry practice of staggering weight releases so cybersecurity teams and safety reviewers can test a model before its weights go public, a process Mistral also followed with its Le Chonk model. Operational urgency for defenders is nil: this is product and industry news.