Inkling 2 Small builds a full city simulator with no pre‑made assets, showcasing its vision and agentic capabilities.
THINKING MACHINES RELEASES INKLING 2 SMALL, 594B MODEL WITH 31 AI INDEX SCORE
*Inkling 2 Small arrives on Nov 5 2026 as a 594 billion‑parameter Mixture‑of‑Experts model. It scores 31 on the Artificial Analysis Intelligence Index and promises long‑horizon autonomous work, but its full weights stay closed until safety testing finishes.*
Thinking Machines Lab unveiled Inkling 2 Small on Nov 5 2026, branding it as the next step in autonomous AI. The model packs 594 billion parameters, but only 17 billion are active during inference, delivering a leaner compute profile while still handling image inputs and variable effort. Its 31 score on the Artificial Analysis Intelligence Index marks a clear leap over the original Inkling, especially in agentic coding, vision, and reasoning. The company is releasing the model through its Tinker platform with closed weights, promising a full Apache 2.0‑licensed release after a week of external safety testing and red‑team review.
What are Inkling 2 Small’s core specifications?
Inkling 2 Small is a Mixture‑of‑Experts architecture with 594 billion total parameters and 17 billion active at inference. It accepts image inputs and supports variable effort, allowing users to trade compute for performance. The model posted a 31 score on the Artificial Analysis Intelligence Index, outpacing the first‑generation Inkling across agentic coding, vision, knowledge work, and reasoning. Training incorporated more efficient pre‑training and a large, unified reinforcement‑learning run, delivering gains at every stage of the pipeline.
How does the model’s training differ from previous generations?
The new generation was trained on hundreds of thousands of long‑horizon coding tasks using multiple agent harnesses and context‑compaction strategies. The synthetic coding data pipeline was upgraded, moving beyond unit‑test rewards to include an agentic RL grader that evaluates problem‑solving approach, not just pass/fail. Early checkpoints, released under the name Fledge‑Alpha, were evaluated on OpenCode to gauge real‑world coding performance. These changes make Inkling 2 Small stronger at coding and easier to embed in autonomous agent loops.
What new capabilities does Inkling 2 Small unlock for developers?
Vision and agentic abilities let the model iteratively build and refine websites, games, and 3D designs with interactive elements. Demonstrations include a Three.js city simulator that procedurally generates every building, texture, and vehicle from a single road, and a Python‑based PC emulator that boots a real Linux kernel in a browser. Inkling 2 Small also produced a hand‑drawn valley with origami cranes in Blender, later improving water and chimney details, and generated floor‑plan walkthroughs from maps. All outputs are available as open videos.
How does Inkling 2 Small perform on benchmark tests?
On the external PMPM‑Hard GPU kernel‑optimization benchmark, Inkling 2 Small recorded an AUC of 494.27, solved a mean of 54.25 out of 69 tasks, and achieved a Pass@4 rate of 64 %. GPT‑6 Astra led the field with an AUC of 616.39 and a Pass@4 of 63 %, while other models trailed further behind. The results place Inkling 2 Small solidly in the upper tier of current large‑language models, confirming its claim of superior long‑horizon autonomous work.
Frequently asked questions
When will Inkling 2 Small’s full model weights be released?
The full weights will be released under an Apache 2.0 license next week after external safety testing and red‑team review.
What are the parameter counts of Inkling 2 Small?
Inkling 2 Small has 594 billion total parameters, with 17 billion active during inference.
How does Inkling 2 Small perform on the PMPM‑Hard benchmark?
It achieved an AUC of 494.27, solved a mean of 54.25 out of 69 tasks, and scored 64 % on Pass@4, placing it behind GPT‑6 Astra but ahead of many peers.
Inkling 2 Small pushes the frontier of long‑horizon autonomous AI, but its cautious rollout underscores the tension between rapid capability gains and responsible deployment. As developers begin to experiment on Tinker, the upcoming open‑weight release will test the community’s ability to harness power without compromising safety. The next week will reveal whether the model’s promise translates into real‑world impact or adds another layer to the ongoing debate over open AI development.
Sources: Hacker News – Inkling 2 Small – https://thinkingmachines.ai/news/inkling-2-small/