Thinking Machines Launches Inkling-Small: 12B Active Params Beats Trillion-Param Model
新智元 · wechat · 2026-07-31
Thinking Machines has officially released its second open-source model, Inkling-Small. Using a MoE architecture, it has 276B total parameters but only activates 12B per inference, supporting a 1M token context. The company achieved a 'student surpassing the teacher' scenario by using the original Inkling for on-policy distillation followed by two weeks of agentic coding RL, setting a new open-source SOTA in reasoning and agentic tasks with a quarter of the original size.
Additionally, the article highlights the company's recent executive turnover: four of its six co-founders have left, including Lilian Weng who returned to OpenAI. Despite losing core team members, the startup has paradoxically accelerated its model release pace.
Related event: Thinking Machines Releases Open-Source Model Inkling-Small(37 posts)→
More from Companies & People
- Sarvam Unveils 105B Model, India's Largest Blackwell Cluster, and New AI Agents — itsOmSarraf_ · 2026-07-31
- Ex-OpenAI Employee Starts Company for High-Quality RL Datasets, Citing Poor LLM Generalization — burny_tech · 2026-07-31
- Palantir CEO Blasts AI Rivals for Selling Tokens, Advocates Value-Based Pricing — eliano · 2026-07-31
- Stone Tao Joins Physical Intelligence to Focus on Robotics Simulation — svlevine · 2026-07-31
- Why Tell Us Now? Trask Questions Timing of AI Labs' Security Disclosures — iamtrask · 2026-07-31
- Gary Marcus Rounds Up Seven Major AI Blunders and Controversies — Gary Marcus · 2026-07-31