Inkling 2 Mini: The Looped Transformer That Puts Frontier-Level Coding on Your Laptop
Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, and it may be the most interesting open model of the year — not because it is the biggest, but because it does more with less. The 26.4-billion-parameter model runs on a laptop, yet its maker says it leads every open-weights model under 100 billion parameters on agentic coding benchmarks. The trick is a looped transformer: instead of adding more layers, the model runs its middle layers three times for every token.
What Inkling 2 Mini Is
Inkling 2 Mini is a general-purpose multimodal model that accepts text and image inputs and generates text outputs, according to its model card on Hugging Face. It is a sparse mixture-of-experts (MoE) design with 26.4 billion total parameters and 3.4 billion active per token, a context window of up to 1 million tokens, and an Apache 2.0 license. The weights are released openly on Hugging Face, with hosted inference and fine-tuning available through the lab’s Tinker platform.
The model follows Inkling 2 Small, a 594-billion-parameter model that Thinking Machines released on November 5. Where Small was built for scale, Mini is built for the machine on your desk. Within a day of release, it was the number one trending model on Hugging Face, with 1,649 likes, according to coverage by Medium’s Sajjad Pakdamansavoji.
The Loop Is the Real Idea
A standard transformer passes each token through every layer once. Inkling 2 Mini stores 24 decoder layers, but it does not use them once each. Layers 0 to 5 run once, layers 6 to 17 run three times with shared weights, and layers 18 to 23 run once — producing 48 layer visits per token from 24 layers of stored weights, as detailed in the Hugging Face model card and in analysis by MarkTechPost.
Each pass through the looped core starts from the same inputs — the output of the first six layers — rather than simply feeding outputs forward. The practical effect, according to The Decoder, is that the model can reason internally for longer without generating more text, finding the right answer with fewer tokens. Because of the loops, the lab says 5.4 billion parameters are effectively active per token, even though only 3.4 billion are active in a single pass.
It is also one of the first open looped-transformer releases, which means the release tests an architectural idea, not just a checkpoint. If looping proves out at this scale, future models may get smarter by thinking longer rather than by growing larger.
Small Enough to Run on a Laptop
The full checkpoint is a single download of about 20 gigabytes, in mixed NVFP4 and MXFP8 precisions, according to the Hugging Face model card. Thinking Machines lists three supported setups: one NVIDIA DGX Spark, Apple silicon devices with 32 gigabytes or more of memory, or other GPUs with 32 gigabytes or more of VRAM. Quantized GGUF builds are available through Unsloth and llama.cpp, and serving support covers Transformers, vLLM, SGLang, Ollama, NVIDIA Dynamo, and TokenSpeed.
To make the point, Thinking Machines researchers ran the model offline on a MacBook Pro with an M5 Pro chip and 48 gigabytes of memory — and had it build a playable flight simulator in a few minutes, according to MarkTechPost and Futurum Group. That is the kind of demo that would have sounded implausible for a model of this capability even a year ago.
How It Performs
All of the benchmark figures below are self-reported by Thinking Machines, and independent verification is still pending — a caveat worth stating plainly. On coding, the model’s strongest area, Inkling 2 Mini scores 79.2 percent on SWE-bench Verified (bash only) and 77.9 percent on SWE-bench Pro v2 (Hard), according to MarkTechPost. It scores 48.2 percent on DeepSWE v1.1 and 46.5 percent on the Vals Index v2.2, according to The Decoder, and 20.1 on the Artificial Analysis Intelligence Index, tied with Qwen3.8-27B.
Hugging Face’s summary of the lab’s comparison table says Mini leads in 32 of 48 benchmarks against four rival open models. Futurum Group reports that the model leads every sub-100-billion-parameter rival on agentic coding benchmarks. The honest reading is that this is a strong open-weights coding model with lab-reported numbers — impressive, but the community will want to see independent runs before treating the rankings as settled.
Why This Matters Beyond One Model
Inkling 2 Mini arrives in a week when machine learning is visibly pushing into new territory. On November 20, Anthropic mathematician Levent Alpöge announced that he and Claude had found a proof of the Riemann Hypothesis, formalized in 17 million lines of Lean code, according to CNN — a claim that has not been independently verified and remains listed as unsolved by the Clay Mathematics Institute. The same day, NYU mathematician Tristan Buckmaster and Alpöge posted a 79-page paper claiming a finite-time blowup proof for the three-dimensional Navier-Stokes equations, with a 600,000-line Lean 4 formalization built with heavy assistance from Anthropic and OpenAI models, according to Emergent — though the critical numerical certificate was not released, only a cryptographic hash promising future availability.
Both claims are unverified, and both may yet fall apart under scrutiny. But together with Inkling 2 Mini, they point in the same direction: machine learning is no longer just a tool for writing code or answering questions. It is becoming a collaborator in research itself — and, with looped architectures like Inkling 2 Mini, a collaborator that can increasingly run on the computer you already own.
The Bottom Line
Inkling 2 Mini is a bet that the next leap in machine learning will come from using parameters more cleverly, not from adding more of them. A 26.4-billion-parameter model that loops its middle layers, fits in about 20 gigabytes, runs on a 32-gigabyte Mac, and posts leading open-weights coding scores is a genuine step toward capable AI that lives on personal hardware rather than in a data center. The benchmarks are self-reported and need independent confirmation — but the architecture is open, the weights are free under Apache 2.0, and anyone with a modern laptop can now test the claim for themselves.
Sources: Hugging Face model card for thinkingmachines/Inkling-2-Mini; MarkTechPost, “Meet Thinking Machines Inkling 2 Mini” (November 21, 2026); The Decoder, “New open model Inkling 2 Mini thinks in loops to solve problems with fewer words and still runs on a laptop”; Futurum Group, “Open-Weight Efficiency: Inkling 2 Mini” (November 21, 2026); Medium, Sajjad Pakdamansavoji, “Thinking Machines’ Inkling 2 Mini Loops Its Middle Layers for Reasoning, Not Knowledge”; CNN, “Anthropic worker claims to have solved a Holy Grail of math problems with the help of AI” (November 21, 2026); Emergent, “AI Models Help Prove Navier-Stokes Blows Up in Finite Time” (November 21, 2026).
Edited by Muse @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous
Never miss a customer call again — MAJ.COM’s AI receptionist answers every call, texts customers back instantly, and books the appointment for you. See how it works at https://MAJ.COM
Discover more from QUE.com
Subscribe to get the latest posts sent to your email.
