Thinking Machines Lab introduces Inkling 2 Small with 594B parameters

Thinking Machines Lab introduced Inkling 2 Small on Nov 5, 2026. The company describes it as a Mixture-of-Experts model with 594B total parameters, 17B active parameters, image inputs and variable effort. It says the model scores 31 on the Artificial Analysis Intelligence Index, surpassing its previous models.

Availability and weights

Inkling 2 Small will initially be available through Tinker with closed weights, following the approach the company outlined in a safe path to open weights. Thinking Machines Lab says it intends to make the full weights available under an Apache 2.0 license next week after external safety testing and red-teaming. The model can be tried on Tinker Playground today and is available for fine-tuning and sampling on Tinker. Checkpoints will be downloadable with the release of the weights.

Capabilities and vision

The company says Inkling 2 Small is more capable across the board than its first generation, particularly in long-horizon autonomous work. Its vision and agentic capabilities let it use visual feedback to iteratively build and refine websites, games and 3D designs with interactive elements. Examples in the announcement include a Three.js city simulator with no pre-made assets, a from-scratch PC emulator in Python that boots a real Linux kernel, and a floor-plan walkthrough generated from a map.

Inkling 2 Small processes image inputs with a 292M-parameter vision encoder based on Perception Encoder. It uses 20×20-pixel patches with 2×2 spatial pooling and interleaves image patches with text tokens. The company reports strong visual capabilities across OCR, grounding, infographic interpretation and visual reasoning, with scores including 81.6% on MMMU Pro (Standard 10), 64.7% on PerceptionBench and 83.8/91.5% on ScreenSpot Pro original and with Python.

Coding and autoresearch

Thinking Machines Lab says improvements at every training step, including more efficient pre-training and a large unified RL run, contributed to the model. Inkling 2 Small was trained with RL on hundreds of thousands of long-horizon coding tasks across multiple agent harnesses and context-compaction strategies. The company increased the quality of its synthetic coding data pipeline and moved beyond unit-test rewards alone; many tasks included an agentic RL grader that evaluated how the model approached a problem, not just whether its code passed. Early checkpoints were tested on OpenCode under the name Fledge-Alpha.

The model was also trained on long-horizon autoresearch tasks. In one test, it autonomously trained Qwen3.5-4B on Tinker to predict time series such as weather and Wikipedia page views. After 4.5 hours of work and $426 in Tinker spend, the trained model had more than doubled its forecasting score.

Knowledge work

Thinking Machines Lab says Inkling 2 Small accelerates knowledge work in realistic professional environments, looking up information, calling tools over MCP and producing deliverables. It reports an ELO score of 1527 on AA-Briefcase, 70.5% on JobBench, 73.9% on Vals Legal Index and 71.8% on EnterpriseBench-CoreCraft. The company says the model is especially strong in finance and quantitative reasoning over spreadsheets, filings and documents, and that among open weight models it is the best performing model across several finance benchmarks.

Sources: Thinking Machines Lab · Techmeme