Inkling 2 Small, announced by Thinking Machines on November 5, 2026, is a mixture-of-experts model with 594B total parameters and 17B active. The announcement comes with long-horizon coding, research and visual demos and a long benchmark list. This article is less about ranking and more about how to read those numbers and how to check them on your own code. Every figure below comes from the announcement; I did not reproduce any of them.
What was released
- Architecture: MoE, 594B total and 17B active. It accepts image input, supports variable effort, and has no audio.
- Availability: initially through Tinker (playground, fine-tuning, sampling) with closed weights. The post says full weights are planned under Apache 2.0 "next week," after external safety testing and red-teaming. That is a plan; wait for the actual release before relying on it.
- Context: no single maximum is stated. Pretraining used 512K and 1M phases, and RL used 256K with compaction.
- Training: 8.4T tokens pretrained natively in NVFP4, then further phases; the RL run was 1,075 steps on 2,048 GPUs over 9 days.
Coding-related numbers
| Benchmark | Reported | Label in the post |
|---|---|---|
| DeepSWE v1.1 | 72.1% | 2nd |
| SWEBench Pro v2 (Hard) | 86.3% | 2nd |
| FrontierCode v2.0 | 25.4% | 2nd |
| Vibe Code Bench v1.1 | 81.6% | no label |
| Terminal Bench Science v0.1 | 5.7% | no label |
| PatchEval | 79.6% | 2nd |
"2nd" means second within the comparison set the post chose. I could not independently verify the full list of compared models or whether each was measured under the same conditions.
The conditions are part of the number
All evaluations ran at effort 0.99 (the post calls it xhigh), temperature 1.0, and a maximum trajectory of 1M tokens, up from 256K for the earlier Inkling Small. These scores are closer to an upper bound with a generous token budget than a guarantee at smaller budgets or lower effort. The post lists no pricing, so cost and latency are unknown.
Weaknesses the announcement admits
- It is not succinct and sometimes picks the wrong tool, especially in codebases with many skills.
- It is less tuned for chat and overreacts to some safety-related queries.
- It trails closed-weights models such as GPT-6.1 Sol on the post's alignment evaluations.
The first point matters for agent harnesses. If tool choice degrades as the toolbox grows, start with a short tool list. That is my practical inference from the stated weakness, not a tested result.
Comparing on your repo
Instead of leaning on public benchmarks, try the steps below. They are a proposal; I have not run this experiment.
- Pick 10 to 20 recently resolved issues and pin the pre-fix commit and the failing test.
- Give your current model and Inkling 2 Small the same prompt, tools and budget (tokens, time).
- Record success rate, human fix-ups, unneeded file changes, tokens, time and cost.
- Rerun at a different effort level to see sensitivity.
For the basics of reading scores, see how to read AI benchmark scores and why scores change when a benchmark version changes.
Conclusion
Inkling 2 Small is notable as a large MoE coding model slated for open weights. What is confirmed today is the publisher's own measurements and a release plan. Check that the weights and license actually ship, and decide on adoption only after comparing success rate and fix-up counts on your own tests.
Sources
- Thinking Machines, Introducing Inkling 2 Small (2026-11-05, checked 2026-11-07)