
A frontier MoE model for agentic coding and autoresearch
Inkling 2 Small is Thinking Machines Lab's second-generation AI model, a Mixture-of-Experts model with 594B total parameters (17B active) that supports image inputs and variable effort. It excels at long-horizon agentic coding, vision, knowledge work, and autonomous research tasks like GPU kernel optimization and autonomous model training. It is initially available through Tinker with closed weights, with full weights planned to be released under an Apache 2.0 license after external safety testing.
Inkling 2 Small is a MoE model with 594B total parameters and 17B active parameters, making it a large but sparsely activated model.
Supports image inputs, enabling capabilities like incorporating visual feedback to iteratively build and refine websites, games, and 3D designs with interactive elements.
The model supports variable effort, letting users adjust how much reasoning the model applies.
Trained with RL on hundreds of thousands of long-horizon coding tasks across multiple agent harnesses and context-compaction strategies, including tasks graded by an agentic RL grader that evaluated how the model approached problems, not just whether code passed unit tests. Scores 86.3% on SWEBench
Trained on long-horizon autoresearch tasks including agentic GPU kernel optimization (scoring 54.25/69 mean tasks solved on PMPP-Hard) and ML tasks such as autonomously post-training a model across a range of GPU architectures.
Available for fine-tuning and sampling on Tinker today, with checkpoints downloadable once the full weights are released.
Initially available with closed weights through Tinker, with full weights intended to be made available under an Apache 2.0 license one week after launch, following external safety testing and red-teaming.
The increase in capabilities comes from improvements at every step of the training process, including more efficient pre-training and a large, unified RL run covering long-horizon coding and autoresearch tasks.
Users can try the model on Tinker Playground today, and use it for fine-tuning and sampling on Tinker; checkpoints become downloadable when the weights are released.
It is the second model generation from Thinking Machines Lab: a Mixture-of-Experts model with 594B total parameters and 17B active parameters, supporting image inputs and variable effort.
At launch, it was available through Tinker, including the Tinker Playground for trying it and Tinker for fine-tuning and sampling. Checkpoints become downloadable when the weights are released.
Not at launch. The model initially ships with closed weights through Tinker, and Thinking Machines Lab stated it intends to release the full weights under an Apache 2.0 license the following week, after external safety testing and red-teaming.
It scored 31 on the Artificial Analysis Intelligence Index, surpassing the company's previous models. On coding benchmarks it achieved 86.3% on SWEBench Pro v2 (Hard), 72.1% on DeepSWE v1.1, and 81.6% on Vibe Code Bench v1.1, generally second-highest in tested rows. On PMPP-Hard kernel optimization it solved a mean of 54.25 of 69 tasks in 10 hours, and it achieved a 6.02x speedup on DeepSeek-V3.2 sparse attention kernels.
Inkling 2 Small autonomously trained a LoRA of Qwen3.5-4B on Tinker to forecast time series such as weather and Wikipedia page views, writing its own evaluator, experimenting with datasets and SFT/RL recipes, and debugging along the way. After 4.5 hours and $426 in Tinker spend, the model more than doubled its forecasting score.
November 5, 2026.