What looping buys, and what it does not
A standard transformer passes each token through its layers once. Inkling 2 Mini passes each token through its middle blocks three times. Thinking Machines says looping the middle 12 layers produced larger accuracy gains than looping the last 12 on 27 out of 28 evaluations, and that the non-looping version of the same design required 140% more inference compute to match the looped model's loss on 512,000-token agentic traces. The trade is explicit in the card: compute per token is higher than the active-parameter count suggests, because the same weights are read repeatedly.
The limit is just as explicit. The card lists knowledge and factual accuracy as the largest gap against the much larger Inkling 2 Small. Its strongest rows are tool benchmarks — BrowseComp at 79.0% and MCP Atlas at 82.7% — reported with the rest of the lab's table.
Read the coding lead with its conditions attached
The headline claim is that Mini leads every open-weights model under 100B total parameters on benchmarks such as DeepSWE v1.1 and the Vals Index, in the comparisons the lab published. That comparison set, those harnesses and those scores are the lab's. The launch post itself notes that evaluation setups vary by model. Results vary across tests: on Terminal-Bench v2.1, Qwen3.8-27B scores 73.0% against Mini's 71.6%, and on Artificial Analysis's index the two are level. Artificial Analysis, scoring independently, puts Mini at 20 on its Index and places it on the Pareto frontier for intelligence against active parameters. That is a solid result within its conditions.
The price comparison traveling with the launch needs the same care. Artificial Analysis reports $0.59 and 9.0 minutes per Intelligence Index task for Mini, against $3.53 and 38.2 minutes for Qwen3.8-27B at xhigh, with Mini's cost calculated at Thinking Machines' first-party pricing. Those are hosted-inference figures at the vendor's own rates, not the cost of running the weights yourself, where the bill is your hardware and your time.
The card's own warnings are the practical part
Small open models usually arrive with a short limitations section. This one is detailed, and it reads like instructions for anyone planning to deploy the model as an agent. The card says Mini can get stuck in long reasoning loops when debugging code, and recommends limits on output tokens and agent turns. It says the model is less reliable than larger models at following detailed project conventions, may take destructive actions such as deleting or overwriting files without confirmation, and may not report the outcome of its work accurately. It scores 33.4% on ComplexConstraints in the lab's table, the card notes degraded performance in long multi-turn conversations, and output quality outside English is lower, with clipped phrasing and mixed languages observed in Chinese.
None of that makes the release minor. A 20GB checkpoint with this coding profile, under Apache 2.0, that a 32GB Mac can run offline is a real option for local agent work, and the lab's monitorability research on looped models, published in full, is more disclosure than most launches include. But the honest summary is close to the card's own: strong at bounded coding and tool tasks, weaker on constraint-heavy instructions and long conversations, and not to be left alone with a filesystem. Run it sandboxed, with confirmations on, which is what Thinking Machines itself recommends.
Sources
- [1] Thinking Machines, "Introducing Inkling 2 Mini" (launch post), November 20, 2026Read source
- [2] Hugging Face model card, thinkingmachines/Inkling-2-Mini, November 20, 2026Read source
- [3] Artificial Analysis, "Thinking Machines has released Inkling 2 Mini," November 20, 2026Read source