Thinking Machines releases Inkling 2 Small through Tinker, holding back downloadable weights

Thinking Machines Lab released Inkling 2 Small on November 5, 2026, giving developers access to a model trained for extended coding and research work through Tinker. They can run and fine-tune it now, but cannot download its weights yet. The lab intends to release them next week, after external safety testing and adversarial testing.

A large model with a smaller active slice

Despite its name, Inkling 2 Small has 594 billion parameters—the learned values inside the model—with 17 billion active. It uses a mixture-of-experts architecture, supports image inputs and offers variable effort. The active count describes the portion used, rather than the model’s entire parameter pool.

Thinking Machines reports a score of 31 on the Artificial Analysis Intelligence Index, substantially above its earlier models. Its coding comparisons give a more specific picture: Inkling 2 Small scores 25.4% on FrontierCode v2.0, behind DeepSeek-V4.1-Flash-1017 at 28.5%. On SWE-bench Pro v2’s hard evaluation, it scores 86.3%, below Kimi K3’s 88.2% but above Muse Spark Open’s 82.4%.

Those company-published results put the model between named rivals rather than ahead across every coding test. DeepSWE v1.1 shows the same pattern: Inkling outperforms several models in the comparison, while DeepSeek retains the highest reported score.

DeepSWE v1.1: the company’s reported comparison

0172.1%

Inkling 2 Small

Thinking Machines reports the second-highest score in its DeepSWE v1.1 comparison.

0277.0%

DeepSeek-V4.1-Flash-1017

DeepSeek has the highest reported score in the same company-published comparison.

Training for the work between the answers

The lab says it trained Inkling 2 Small with reinforcement learning—training guided by rewards—on hundreds of thousands of long-running coding tasks. It used multiple agent harnesses, the software environments in which models take actions, and different strategies for compacting accumulated context.

Passing code tests was not the only reward. Many tasks included an AI grader that assessed how the model approached a problem, not just whether the resulting code passed. Thinking Machines says it also tested early checkpoints in OpenCode under the name Fledge-Alpha to assess coding performance outside its training setup.

Image input adds another feedback channel. The company demonstrates the model using visual feedback to build and revise websites, games and interactive 3D designs. Its examples include a Three.js city simulator with procedurally generated buildings, textures and cars, and a Blender scene revised to improve its water and chimney.

A model that trains another model

A separate launch demonstration goes beyond writing software. Thinking Machines gave Inkling 2 Small a task built by Proximal for an internal benchmark: train Qwen3.5-4B through Tinker to forecast time series, including weather and Wikipedia page views. The company describes an autonomous run lasting about 4.5 hours and costing $426 in Tinker spending.

Inside the forecasting demonstration

• The model analyzed the task and wrote its own evaluation script before training.
• It experimented with datasets and training recipes, trying both supervised fine-tuning and reinforcement learning.
• Its trace shows it finding a sampling-seed mistake and changing the evaluation to use distinct seeds.

That example is an internal demonstration, not a general measure of autonomous research reliability. It does show the workflow the lab is targeting: a model managing evaluation, data preparation, training experiments and debugging, rather than producing a single code answer.

For now, access stays inside Tinker: the Playground supports trying the model, while Tinker offers fine-tuning and sampling. The lab says fine-tuned checkpoints will become downloadable when the full weights are released under Apache 2.0. That separates customization available now from possession of the resulting model files later.