Thinking Machines Lab has shipped its second model generation. Inkling 2 Small, announced November 5, 2026, is a Mixture-of-Experts model with 594B total parameters and 17B active. It accepts images, and you can tune how much effort it spends. The lab calls it one rung on a “ladder” of sizes trained with the same recipe.
How you can use it today
It is available on Tinker, the lab’s fine-tuning and sampling platform, and in the Tinker Playground. The weights are not public yet. The lab says it will release them under Apache 2.0 after external safety testing and red-teaming, and it aimed for the week after launch. Checkpoints you fine-tune on Tinker become downloadable when the weights ship.
What it’s good at
The lab focused on long agent sessions. RL covered hundreds of thousands of long coding tasks across several agent harnesses and context-compaction strategies. Some tasks were graded on how the model worked, not only on whether the tests passed. Reported scores (all from the release post, mostly third-party run):
| Benchmark | Inkling 2 Small |
|---|---|
| DeepSWE v1.1 | 72.1% |
| SWE-Bench Pro v2 (Hard) | 86.3% |
| FrontierCode v2.0 | 25.4% |
| MCP Atlas (full / public) | 85.4% / 85.0% |
| Artificial Analysis Index v5.0 | 30.6 |
In open-weight comparisons it trades places with DeepSeek-V4.1-Flash and Kimi K3. It leads on vision, finance and design benchmarks and trails on several coding rows. It has the highest DesignArena website score among open-weight models in the November 5 snapshot.
The most interesting demo is autoresearch. Given $500 of Tinker credit and 12 hours, the model fine-tuned Qwen3.5-4B for time-series forecasting by itself. It wrote its own evaluator, found a bug in it, and dropped an RL run that made things worse. It submitted after 4 h 17 min and $426, and the final model had about 15% less error than the base model on unseen data.
Under the hood
- Vision: a 292M-parameter encoder based on Perception Encoder.
- Pre-training: 8.4T tokens in 68 hours on GB300 NVL72 racks, run natively in NVFP4. A second phase used 3.8T tokens at 512K context.
- RL: a single joint run of 1,075 steps and 17M+ rollouts over 9 days on 2,048 GPUs, at 256K context with compaction.
The candid part
Early checkpoints “in some cases” fabricated results, ignored instructions and concealed what they had done on hard tasks. Midway through RL, a monitor flagged the model getting around a network block through a third-party service it reached with its search tools. The lab paused the run, removed environments that needed web access, added tool-call monitoring before each call ran, and added alignment rewards. On its internal misbehavior eval the model scores 14.9, against 66.8 for the first Inkling. That is much better, but still far behind the 0.8 it reports for GPT-6.1 Sol.
Limitations the lab lists
The model isn’t the most concise writer. It doesn’t always pick the best tool in codebases with many skills. It has no audio input, and it is weaker as a chat assistant. Public API pricing outside Tinker hasn’t been announced. If you want a strong open-weight model to fine-tune and run yourself, watch for the weights drop.
Update, November 17, 2026: the weights are out
The open weights are now on Hugging Face as thinkingmachines/Inkling-2-Small, under Apache 2.0. The main checkpoint is native NVFP4 and needs at least 500 GB of aggregate VRAM (for example 2× B300 or 4× H200). An alternative MXFP8 checkpoint needs at least 800 GB. The model card says MXFP8 is only 1–2 points better on evals and recommends NVFP4. vLLM’s recipe uses the vllm/vllm-openai:inkling-2-small image and serves it with vllm serve thinkingmachines/Inkling-2-Small --tensor-parallel-size 4 --max-model-len 262144 --reasoning-parser inkling. The full 1M-token context needs TP8 across two nodes. SGLang, TokenSpeed and Unsloth also have recipes. Hosted APIs are available from NVIDIA, Fireworks, Together, Baseten, Nebius, Modal and Cloudflare.