The Pulse
Thinking Machines Lab Plans Apache 2.0 Weights for Inkling 2 Small, a 594B-Parameter Model
Thinking Machines Lab introduced Inkling 2 Small, a Mixture-of-Experts model with 594B total parameters, image inputs and variable effort. The company says it will initially offer the model through Tinker with closed weights, and intends to
AI.info Team · Written with AI · fact-checked ·
Thinking Machines Lab introduced Inkling 2 Small on Nov. 5 as its second model generation. The Mixture-of-Experts model has 594B total parameters and 17B active parameters, supports image inputs and offers variable effort. The company says the model will initially be available through Tinker with closed weights. It intends to release the full weights under an Apache 2.0 license next week, after external safety testing and red-teaming.
Users can try Inkling 2 Small in Tinker Playground, and use it for fine-tuning and sampling on Tinker. The company says downloadable checkpoints will arrive with the release of the model weights.
Thinking Machines says it trained Inkling 2 Small with reinforcement learning on long-horizon coding tasks
Thinking Machines reports a score of 31 for Inkling 2 Small on the Artificial Analysis Intelligence Index, which it says far surpasses the scores of its previous models. The company attributes the gains to improvements at every step of its training process, including more efficient pre-training and a large, unified reinforcement-learning run. It says it trained the model with reinforcement learning on hundreds of thousands of long-horizon coding tasks. Those tasks ran across multiple agent harnesses and context-compaction strategies.
The company also says it increased the quality of its synthetic coding data pipeline and moved beyond unit-test rewards alone. Many tasks used an agentic reinforcement-learning grader to assess the model’s approach as well as whether its code passed. Thinking Machines tested early checkpoints on OpenCode under the name Fledge-Alpha. It says the model follows instructions better than Inkling, considers changes carefully and verifies its work without being asked.
Company-reported coding benchmark scores
Inkling 2 Small scored 25.4% on FrontierCode v2.0, compared with 28.5% for DeepSeek-V4.1-Flash-1017 and 17.3% for Kimi K3. On DeepSWE v1.1, it scored 72.1%, below DeepSeek-V4.1-Flash-1017 at 77.0% and above Kimi K3 at 68.5%. Its SWEBench Pro v2 (Hard) score was 86.3%, and Kimi K3 scored 88.2%. On Vibe Code Bench v1.1, Inkling 2 Small scored 81.6%; DeepSeek-V4.1-Flash-1017 scored 87.7%.
The company also reports results from PMPP-Hard, an external GPU kernel-optimization benchmark adapted for cross-model evaluation. Inkling 2 Small solved a mean of 54.25 tasks out of 69 over four independent attempts, with a 10-hour budget. Its reported area under the mean curve was 494.27 task-hours. In the evaluation, each model uses its maximum available reasoning setting, the PI harness and a shared compaction policy.
Autoresearch runs include a $426 forecasting test
On a sparse-attention kernel-optimization task, Thinking Machines reports a preliminary 6.02× speedup over the original kernels for Inkling 2 Small. The company says the speedups are for those kernels alone, not end-to-end training. In a separate time-series forecasting task, the model trained Qwen3.5-4B on Tinker using data from electricity demand, wind and solar power, weather, air quality and Wikipedia page views. The task included analyzing the data, writing an evaluator, experimenting with training recipes and debugging the process.
Thinking Machines says Inkling 2 Small stopped after 4 hours and 17 minutes, having spent $426 of a $500 budget. Its best submission achieved a score of 0.289, compared with 0.191 for the base Qwen3.5-4B model. On the task’s unseen August–September data, the final model made about 15% less error than the base model.
Thinking Machines says image input supports visual creation and evaluation
Inkling 2 Small processes images with a 292M-parameter vision encoder based on Perception Encoder. The company describes visual uses that include building and refining websites, games and 3D designs with interactive elements. Its examples include a city simulator built with Three.js, a Linux PC emulator written in Python and a floor-plan walkthrough generated from a map.
On the November 5 DesignArena Website leaderboard snapshot, Inkling 2 Small scored 1,340, behind Muse Wonder at 1,366 and Muse Spark 1.3 at 1,351. The company also says it tested the model against some of its researchers in a wine-tasting event. It says the model outperformed them using vision alone, correctly identifying three wines from a single picture in 19 out of 20 trials.