Confirmed 20 Nov 2026 · Labs A one-minute read
Thinking Machines released a 26.4 billion parameter open-weights model that runs on a laptop.
Inkling 2 Mini loops its middle layers 3 times per token, so 24 stored layers do 48 layers of work in about 20GB. The lab says looping adds no knowledge.
Small models get better by shrinking big ones or training them longer.
Depth can be rented at run time by reusing the same layers.
Why it matters
Looping buys reasoning depth without storing more weights, which is why a laptop can run it. The lab also reports the model reasoning inside the loops rather than in written text. If that holds at larger scale, the written reasoning that safety monitors read gets thinner.
Established
- Open weights are published on Hugging Face: 26.4 billion total, 3.4 billion active.
- It scored 48.2% on DeepSWE v1.1 against Qwen3.8-27B's 42.2%.
- Two middle blocks run 3 times, giving 48 layer passes from 24 layers.
Not established
- That it beats larger models. It trails them on CritPt and HLE-Diamond.
- That looping hides reasoning. The lab found no meaningful loss of monitorability.
- Its licence terms. The lab's own page says only open weights.
The five-minute readThe record, the timeline, the weight and 3 sources
The record
Thinking Machines Lab's post, dated November 20, 2026, introduces Inkling 2 Mini as open weights "small enough to run on your laptop". Architecture: a mixture-of-experts looped transformer, 26.4 billion total parameters, 3.4 billion active per token, 5.4 billion weight activations once looping counts. 24 layers sit in 4 blocks of 6. The first and last blocks run once; blocks 2 and 3 share weights across 3 passes, giving 48 layer applications per token from 24 stored layers, each pass reading a message from the previous one. Looping the middle layers beat looping the last layers on 27 of 28 evaluations. Memory footprint is about 20GB. Most weights are NVFP4, the last 4 layers MXFP8, deployed in the precision they were trained in, so the lab claims no conversion loss. Training started with 384 experts choosing 8 per token; at midtraining two-thirds were pruned, keeping the 128 most used per layer, then healed on 1 trillion tokens distilled from a larger internal Inkling model, then supervised fine-tuning and reinforcement learning totalling 2.2 trillion tokens over 2,000 steps. Weights are on Hugging Face as thinkingmachines/Inkling-2-Mini, with Tinker, Ollama for Macs with 32GB or more, and GGUF builds via Unsloth and llama.cpp. Against Qwen3.8-27B: DeepSWE v1.1 48.2% to 42.2%; SWE-bench Pro v2 hard 77.9% to 73.8%; SWE-bench Verified bash only 79.2% to 77.8%; MCP Atlas 82.7% to 77.3%; BrowseComp 79.0% to 61.6%; Artificial Analysis Index v5.0 20.1, a tie. Qwen leads Terminal Bench v2.1, 73.0% to 71.6%, and ARC-AGI v1. On Stanford's intelligence per joule measure on an NVIDIA DGX Spark, Inkling scored 0.703 to Qwen3.6-35B-A3B's 0.430 and Qwen3.5-27B's 0.036, at a different precision. The lab's own limits: knowledge "is mostly a function of total parameter count", so looping adds none; the model "is not the strongest small model for multi-turn chat or for instructions with many constraints", at 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1; and on monitorability it concludes that "at this scale we do not see a meaningful reduction in the monitorability of their reasoning". The no-chain-of-thought time horizon rose from 16 seconds at one pass to 37 seconds at 3, measured on a separate variable-loop research model trained only at 3 passes. Not stated on the lab's page: the licence, the context window, and any price.
How it got here
- Sep 22
Xiaomi released the highest scoring open weights model in the world, carded in Edition 031. Open weights had become a place where phone makers and recruitment firms competed, not only laboratories.
- Oct 25
A Chinese recruitment firm's lab published a 33 billion parameter coding model free, carded in Edition 062. The small open model lane was already crowded on parameter count.
- Nov 18
Google shipped Gemini 4 Neon and OpenAI GPT-6.2 Astra, the second rated High for improving AI itself in Edition 100. Both are closed, large, and sold by the call.
- Nov 20
Thinking Machines publishes Inkling 2 Mini's weights, its full benchmark tables, its efficiency figures and its own monitorability results, including the places it loses to Qwen3.8-27B.
Qwen's team, named as the comparison in every table; Google and NVIDIA, whose Gemma and Nemotron models sit far below in the same rows; Apple, since the deployment target is a Mac with 32GB; safety researchers who monitor written reasoning for a living; every company paying by the token for work a laptop may now do.
What to watch
- Would strengthen itA second laboratory ships a looped open-weights model, or a looped model under 50 billion parameters tops the Artificial Analysis index for its size class, by March 31, 2027.
- Would weaken itIndependent evaluations place Inkling 2 Mini below Qwen3.8-27B on a majority of public coding benchmarks, or the efficiency gap closes at matched precision, by February 28, 2027.
- Would strengthen itThinking Machines publishes a looped model above 100 billion parameters, or names the size complaint about Inkling 2 Small as the reason for this release, by June 30, 2027.
Why it weighs this much
Cost is high: 1 trillion tokens of distillation plus 2.2 trillion tokens of reinforcement learning over 2,000 steps, and the lab gave the result away. Commitment is high, because published weights with full benchmark tables, including the rows it loses, invite replication by anyone. Reversibility is low in the strict sense: weights on Hugging Face and in GGUF builds cannot be recalled. Persistence is structural. A trick that trades memory for repeated passes applies to every model that has to fit on a device, so this outlives the particular release.
The words
- Looped transformer
- A model that runs the same layers more than once per token instead of stacking new ones. It buys depth with time rather than memory, so the file stays small.
- Open weights
- The trained parameters are published, so anyone can run or fine-tune the model on their own machine. It is not the same as publishing the training data or code.
- Active parameters
- In a mixture-of-experts model, the fraction of parameters used for any one token. It sets the speed and cost, while total parameters set the memory and the knowledge.
- Monitorability
- How much of a model's reasoning appears as readable text that a safety system can inspect. Reasoning done silently inside layers cannot be read.
Sources
- TThinking Machines Lab: Introducing Inkling 2 Mini, architecture, benchmarks, training and monitorability results, Nov 20, 2026thinkingmachines.ai
- TThe Next Web: Thinking Machines releases Inkling 2 Mini, a looped AI model for laptops, Nov 21, 2026thenextweb.com
- TTestingCatalog: Thinking Machines releases Inkling 2 Mini open-weight modeltestingcatalog.com
Your call
Which explanation do you believe?
A call is final. Other people's calls show after yours.