AI CURATED HOT STORY · 11/6/2026
Thinking Machines has released its second-generation model, Inkling 2 Small, which scores 31 on the Artificial Analysis Intelligence Index v5.0.1, an 18-point improvement over its predecessor and on par with Gemini 3.8 Flash (high), while surpassing Xiaomi's MiMo-V2.6-Pro. The model uses a sparse Mixture-of-Experts (MoE) architecture with 594B total parameters but only 17B activated, supports image input and variable reasoning intensity, balancing efficiency and performance.
Technically, Inkling 2 Small shows significant gains in agentic tasks: AutomationBench-AA jumps from 5% to 46%, FrontierCode 2.0 from 5% to 25%, and HLE-Diamond from 12% to 29%. Coding ability is mixed: FrontierCode 2.0 scores 25% versus Claude Sonnet 5.5's 11%, but Terminal-Bench 4.0 is only 13% versus 30%. The model is initially available as closed weights via the Tinker platform, priced at $0.30/$1.00 per million input/output tokens and $0.03 for cached input. Officials plan to open full weights under Apache 2.0 next week after external safety testing and red-teaming.
If the open-weight promise is fulfilled, Inkling 2 Small will become the highest-scoring open-weight model outside China, significantly raising the capability ceiling for open models. Its sparse MoE design achieves high performance with very low activated parameters, offering a new option for efficient inference deployment, while the leap in agentic tasks reflects the effectiveness of large-scale unified reinforcement learning. This move could intensify competition between open and closed models, pushing the industry to balance efficiency and safety.