Thinking Machines Launches Inkling 2 Small: 594B-Parameter Second-Generation Model, Open Weights Next Week

What happened

The company announced Inkling 2 Small on Nov 5 on its official blog and LinkedIn. It is Thinking Machines Lab's second-generation model, following the first-generation Inkling released in July: the original Inkling has 975 billion total parameters and 41 billion active, with weights already open; the lighter Inkling-Small has 276 billion total parameters and 12 billion active.

The new model is a mixture-of-experts system with 594 billion total parameters and 17 billion active per token, supporting image inputs and variable thinking effort. The weights are closed at launch; Tinker provides inference, fine-tuning and sampling, and Tinker Playground offers inference access. The company says it plans to release the full weights next week under Apache 2.0, once external safety testing and red-teaming are complete, with fine-tuned checkpoints downloadable alongside the weights.

On capabilities, the company says the new model is stronger across the board than the previous generation in long-horizon autonomous work, agentic coding, vision and knowledge work, with reinforcement learning on hundreds of thousands of long-horizon coding tasks during training. The company also showed an autonomous-research demo: the model trained Qwen3.5-4B on Tinker for time-series forecasting on its own, writing its own evaluator, trying data recipes and training methods and debugging the process; after 4.5 hours and $426 of spend, the forecasting score more than doubled.

Key facts

1. Publisher and date: Thinking Machines Lab, Nov 5, 2026; the company's second-generation model
2. Scale: a mixture-of-experts model with 594 billion total parameters and 17 billion active per token; supports image inputs and variable thinking effort
3. Previous generation: the original Inkling has 975 billion total parameters and 41 billion active, with weights already open; Inkling-Small has 276 billion total parameters and 12 billion active
4. License and access: closed weights at launch, available only through Tinker and Tinker Playground; full weights planned for release next week under Apache 2.0, pending completion of external safety testing and red-teaming
5. Pricing: on Tinker, $0.30 per million input tokens and $1.00 per million output tokens, with cached input at $0.03
6. Benchmarks (company-reported): 31 on the Artificial Analysis Intelligence Index, 18 points above the first-generation Inkling Small; 72.1% on DeepSWE v1.1; 86.3% on SWE-bench Pro v2 Hard; 25.4% on FrontierCode v2.0; 81.6% on MMMU Pro
7. Autonomous-research demo: the model trained Qwen3.5-4B for time-series forecasting on its own in 4.5 hours for $426 of Tinker spend, more than doubling the forecasting score
8. Caveat: the weights are not yet released, and the release date depends on external safety testing and red-teaming being completed; benchmark scores are company-reported with no independent re-run yet