Thinking Machines releases Inkling 2 Small, a 594B MoE model
Thinking Machines Lab released Inkling 2 Small on November 5, 2026, the first model of its second generation. It is a Mixture-of-Experts model with 594B total parameters and 17B active per token, accepts image input and supports variable reasoning effort. You can try it today in the Tinker Playground; the lab says the weights will follow under an Apache 2.0 license next week.
What’s new
- Architecture: Mixture-of-Experts, 594B total parameters, 17B active. It takes images as input. It does not support audio.
- Benchmark score: the lab reports a score of 31 on the Artificial Analysis Intelligence Index, well above its previous models.
- Training: the lab credits more efficient pre-training and “a large, unified RL run”. The coding part used hundreds of thousands of long-horizon tasks across several agent harnesses, graded not only by unit tests but also by an agentic grader that judged how the model approached the problem.
- Real-world testing: early checkpoints ran on OpenCode under the name Fledge-Alpha before launch.
- Autoresearch: the model was also trained on GPU kernel optimization and ML tasks. In one demo it trained a Qwen3.5-4B LoRA on Tinker to forecast time series; after 4.5 hours and $426 of Tinker spend, the lab says the forecasting score more than doubled.
Availability
Inkling 2 Small launches on Tinker only, with closed weights, for inference and fine-tuning. Thinking Machines says it will publish the full weights under Apache 2.0 next week, after external safety testing and red-teaming, and that checkpoints from Tinker fine-tuning will become downloadable at that point. More platforms are promised “in the coming weeks”. The lab frames this staged release as its “safe path to open weights”.
What it means for developers
The model is aimed at agent loops. The lab says it follows instructions better than the first Inkling, considers changes carefully and verifies its work without being asked. Its vision and coding skills let it use visual feedback to refine websites, games and 3D scenes, and the launch demos include a Three.js city simulator and a from-scratch PC emulator in Python that boots a Linux kernel.
If you want to try it now, use Tinker. If you want to self-host, wait for the Apache 2.0 release and check the model card for sizes and formats before planning hardware: at 594B total parameters, the weights are large even though only 17B are active per token.
The lab lists limitations you should know about. The model is not the most succinct communicator, doesn’t always pick the best tool in codebases with many skills, and is less tuned for chat: it can organize content badly, overuse slashes and overreact to mental health and other safety-related questions.
How it compares
All numbers below are from Thinking Machines’ own comparison tables, so treat them as vendor-reported.
| Benchmark | Inkling 2 Small | Best score in its comparison row |
|---|---|---|
| FrontierCode v2.0 | 25.4% | 28.5% (DeepSeek-V4.1-Flash-1017) |
| DeepSWE v1.1 | 72.1% | 77.0% (DeepSeek-V4.1-Flash-1017) |
| SWEBench Pro v2 (Hard) | 86.3% | 88.2% (Kimi K3) |
On the PMPP-Hard GPU kernel benchmark it solved a mean of 54.25 of 69 tasks in a 10-hour budget, behind GPT-6 Astra (63), Kimi K3.1 Pro (60.5) and Claude Opus 5.5 (60). In the lab’s DeepSeek sparse-attention kernel test it reached a 6.02x speedup, against 18.80x for Claude Opus 5.5 and 10.26x for GPT-6 Astra. In short, it is competitive among models with this kind of active-parameter budget on coding, but not the top of the list on kernel work.
For other recent model news, see our post on GPT-6 Cyber.