Thinking Machines Lab released Inkling 2 Mini on November 20, and I read the official announcement and the model card the same day. The headline is simple: a 26.4-billion-parameter open-weights model that runs on a laptop and beats every rival under 100 billion parameters on agentic coding benchmarks. The trick is not more parameters — it is reusing the ones it has. Here is what I found, what the numbers say, and where I would still be careful.
The loop trick, explained simply
Inkling 2 Mini is a looped transformer. It has 24 layers, and the middle 12 run three times for every token the model generates. Thinking Machines says that repeating the middle layers worked better than repeating the final ones on 27 of 28 evaluations, so the design is deliberate, not a gimmick.
The model is a sparse Mixture-of-Experts design: 26.4 billion total parameters, but only 3.4 billion active per token. Looping pushes that to 5.4 billion weight activations per token — more reasoning per token without a bigger model. Most weights are stored in NVFP4 precision, with the last four layers in MXFP8, which is how the whole checkpoint fits in about 20GB of memory.
That footprint is the real story. The official hardware targets are Apple silicon machines with at least 32GB of unified memory, a single NVIDIA DGX Spark, or any setup with 32GB or more of VRAM. Thinking Machines demonstrated it by running the model offline on a MacBook Pro during a flight, where it built a playable flight simulator in minutes. A frontier-class coding agent that works with no internet connection and no cloud bill is a genuinely different proposition.
The benchmarks I checked
All scores below are Thinking Machines' own reported numbers, from the announcement and model card. The coding results are the strongest part:
- DeepSWE v1.1: 48.2% — ahead of every open-weights model under 100B total parameters, and 6 points clear of Qwen3.8-27B at 42.2%.
- SWE-bench Verified: 79.2%, and SWE-bench Pro v2 Hard: 77.9%.
- Vals Index v2.2: 46.5%, again the best reported score among sub-100B open models.
- Artificial Analysis Intelligence Index v5.0: 20.1 — tied with Qwen3.8-27B, not ahead of it.
- Agentic search and tool use: 79.0% on BrowseComp, 85.3% on DeepSearchQA and 82.7% on MCP Atlas.
- Reasoning: 86.6% on GPQA Diamond and 49.3% on ARC-AGI-2.
The honest context: Qwen still wins several rows. On Terminal-Bench 2.1, Inkling scores 71.6% against Qwen3.8-27B's 73.0%. And the weak spots are real — 15.2% on HLE-Diamond, 50.0% on MultiChallenge, and a below-par factuality score on AA-Omniscience. This is a model tuned for one job — long-horizon agentic coding and tool use — and its knowledge recall and multi-turn instruction following lag behind. That matches what the announcement itself says: Inkling 2 Mini "does best on long-horizon agentic work."
For scale, compare it with the other big open release of the month: Alibaba's Qwen4-Max, a far larger flagship. Inkling is not trying to be that. It is trying to be the model you run locally — and on the coding benchmarks that matter for a local agent, it currently leads its size class.
Open weights, and a research sibling
Inkling 2 Mini is released with open weights under the Apache 2.0 licence, available on Hugging Face and through Thinking Machines' managed Tinker platform for inference and customisation. It takes text and image input, and the context window runs to about 1 million tokens.
Thinking Machines also released a companion research checkpoint, TML-d24-varloop-research, which exposes the loop count as an inference-time knob — 1 to 5 passes — so researchers can study what looping does. The safety finding is worth noting: the model's score on agentic-sabotage tests did not trend higher as loops were added, and the model card warns the research variant may behave unpredictably at pass counts other than 3. It is explicitly not for production.
Why this matters — my take
The open-weights race in 2026 has been about efficiency, not just size. Inkling 2 Mini is the clearest example yet: instead of scaling parameters, it spends more compute per token by looping, and gets flagship-level agentic coding out of a 20GB checkpoint. If the benchmarks hold up independently, this becomes a default local coding agent — the model you point at a repo on a laptop and trust to inspect files, run tests, read failures and try again, offline.
That matters because the mid-tier is where AI becomes infrastructure. Google's Gemini 4 Neon and OpenAI's GPT-6.2 Astra and Sol are fighting the same battle from the cloud side — frontier-level coding at mid-tier cost. Inkling attacks from the opposite direction: no cloud at all. The two approaches will meet in the middle, and developers win either way.
One more detail I like: the variable-loop research model. A lab shipping the tool to study its own trick, with the safety results attached, is how open research should work.
The honest limits
- Every benchmark in this article is Thinking Machines-reported. Independent verification is still pending, and vendors always publish their best numbers.
- Knowledge recall is a weak spot: 15.2% on HLE-Diamond and a poor AA-Omniscience score mean this is not the model for trivia-heavy factual work.
- Multi-turn instruction following is middling at 50.0% on MultiChallenge — long conversations may drift.
- The research variable-loop variant is explicitly not for production and may behave unpredictably away from 3 passes.
- The 1M-token context is the advertised maximum; real-world long-context behaviour needs independent testing, just as with Claude Sonnet 4.6 and its 1M-token beta window.
The bottom line
Inkling 2 Mini is the most interesting small model of the month: a looped transformer that punches above its weight on agentic coding, fits in 20GB, and runs on a laptop with no cloud. The open question is whether the coding lead survives independent testing — and whether the weak factuality scores matter for the local-agent jobs it is built for. I will be watching the independent results as they land.
Would you run your coding agent locally on a model like this, or stick with a cloud flagship? Tell me in the comments.
Sources
- Thinking Machines Lab, "Introducing Inkling 2 Mini" (official announcement, November 20, 2026): https://thinkingmachines.ai/news/inkling-2-mini/
- Inkling 2 Mini model card, Hugging Face: https://huggingface.co/thinkingmachines/Inkling-2-Mini


Comments
Post a Comment