Inkling 2 Mini: A 26.4B Open-Weights AI Model That Runs on a Laptop

Thinking Machines Lab has released the small model to go with its big one. Inkling 2 Mini, out on November 20, 2026, is a 26.4-billion-parameter AI model with open weights, a 1-million-token context window and a trick up its sleeve: it reuses its own middle layers three times for every token it generates, squeezing far more work out of a model small enough to run on a laptop. It is the smaller sibling of Inkling 2 Small, the 594-billion-parameter model the company released a week earlier — and where that model needs a GPU cluster, Mini is built for the machine already on your desk.

What Is Inkling 2 Mini?

Inkling 2 Mini is a general-purpose multimodal model: it takes text and images as input and produces text. Thinking Machines designed it for developers building AI-powered applications — coding assistants, agentic and tool-use systems, chatbots and retrieval-augmented generation — as well as for everyday conversational use. It is released with open weights under the Apache 2.0 licence, which means anyone can download, fine-tune and build on it, subject to the company’s Model Acceptable Use Policy.

The headline numbers are striking for a model this size. It has 26.4 billion parameters in total, but only 3.4 billion of them are active for any given token, because it is a mixture-of-experts design: each token is routed to 8 of 128 experts. It accepts images as well as text, and its context window stretches to 1 million tokens — enough to hold a large codebase, a long document set or a very long conversation in a single session.

A developer working at a home office desk with code and an AI assistant panel on a laptop screen

A Looped Transformer: The Same Layers, Three Times Over

The idea that makes Mini different is looping. It is a 24-layer decoder-only transformer, but the middle 12 layers run three times with shared weights, so each token gets 48 layer passes out of just 24 layers of weights. The extra passes add about 1 billion effective active parameters each, bringing the effective active count to roughly 5.4 billion per token.

Looping is a trade: more computation per token without more weights to store. That is exactly what a small local model wants, because the weights are the part that has to fit in memory. The cost is speed — the middle of the model does three times the work — and a quirk Thinking Machines is upfront about: looping lets the model do more computation without reasoning in visible tokens, so there is very little control over its chain of thought, and it was unable to perform covert side tasks in realistic agentic tests. The company studied that monitorability question before release and published its findings alongside the model.

Built to Run on Your Own Hardware

The full checkpoint is about 20GB, shipped in mixed NVFP4 and MXFP8 precision. Thinking Machines officially supports one Nvidia DGX Spark, Apple silicon devices with 32GB or more of memory, and other GPUs with 32GB or more of VRAM. But the community has already pushed it further down the hardware ladder. Unsloth’s GGUF quantisations bring it within reach of ordinary consumer cards: the 2-bit file is just 8.5GB and has been run fully on a 12GB RTX 3060, while the 4-bit file is about 16.4GB. An independent local-run check published the day after release confirms those sizes and notes the catch — long context costs more memory than on a normal 26B model, because each pass through the loop keeps its own KV cache.

Support arrived fast. The model works with Hugging Face Transformers, vLLM, SGLang, Ollama and Unsloth, and llama.cpp merged support for Mini’s looped design on release day, shipping in build b12600. That matters because llama.cpp 1.0 has just given the local AI engine a stable API, and the combination of a capable open model and a stable local runtime is what turns “runs on a laptop” from a demo into something developers can actually ship with.

How It Scores Against Its Rivals

On Thinking Machines’ own evaluations, published on the model card, Mini punches well above its weight — particularly at coding. It scores 48.2% on DeepSWE v1.1 and 77.9% on SWE-bench Pro v2 (Hard), ahead of Qwen3.8-27B’s 42.2% and 73.8% on the same tests, and 79.2% on SWE-bench Verified. On Terminal-Bench v2.1 it reaches 71.6%, and on the agentic search benchmark BrowseComp it scores 79.0%. Its Artificial Analysis Index score of 20.1 ties Qwen3.8-27B exactly — the basis for the claim that it matches 27B-class quality.

Those are the vendor’s numbers, and they should be read as such: independent testing will tell a fuller story. But the pattern is consistent across coding, agentic work and vision — 74.9% on MMMU Pro and 92.8% on OmniDocBench — which suggests this is not a one-benchmark wonder. It is also notably safer-behaved than its rivals on the model card’s own misuse measures: internal misbehaviour rates of 9.8%, against 57.7% for Qwen3.8-27B, though those are the vendor’s internal evaluations.

Close-up of a laptop keyboard with a glowing looping neural network of blue and orange nodes above it

The Honest Limitations

Thinking Machines is unusually candid about where Mini falls short, and the list is worth reading before you hand it real work. It is less capable than Inkling 2 Small across most evaluations, with the largest gaps on factual recall and knowledge-heavy questions, and it is more likely than larger models to state incorrect information with confidence. In internal testing it sometimes got stuck in long reasoning loops or repeatedly re-checked its own work without making progress, particularly when debugging code — the company recommends setting limits on output tokens and agent turns.

The agentic warning is the sharpest one. When using tools, Mini is less reliable than larger models at following detailed instructions and project conventions. It may take destructive actions, such as deleting or overwriting files, without confirmation, and may not report the outcome of its work accurately. The company’s own recommendation is to run agentic workloads in a sandbox and require confirmation before destructive or irreversible actions — advice that applies to every coding agent, but doubly so to a small model running on your own machine. It is also worth noting that local support for looped models is brand new, so speed varies by runtime and hardware, and the model can be over-cautious on sensitive topics.

What It Means for Local AI

Mini lands at a moment when local AI is getting serious. Nvidia’s RTX Spark AI PCs are built to run frontier models on a Windows laptop, Google’s Gemini 4 Neon is pushing frontier-level reasoning into cheaper, faster tiers, and tools like Cowork let several coding agents share one inbox. Mini fits that wave from the open side: a model you can download, inspect, fine-tune and run offline, with no API bill and no data leaving your machine.

That is the real significance. A year ago, a 27B-class model that could code, reason over a million tokens of context and see images was a data-centre workload. Now it is a 20GB download that fits on a gaming laptop. The looped transformer is an elegant way to get there — reuse what you have instead of adding more weights — and if the idea holds up in independent testing, expect to see it copied. For now, Inkling 2 Mini is one of the most capable open models you can actually run at home, with its limitations printed plainly on the box.

Conclusion

Inkling 2 Mini is a small model with a clever idea and an honest spec sheet. The looped architecture gives it the reasoning depth of a much larger model, the open Apache 2.0 licence gives developers real freedom to use it, and the 16GB-friendly quantisations make local AI genuinely practical rather than merely possible. It will not replace a frontier model for the hardest work, and its own makers warn you to keep it sandboxed and supervised. But as a coding assistant, a research helper or a private AI that lives on your own hardware, it is one of the strongest open releases of the year — and a sign that the most interesting AI of 2026 may not be the biggest model, but the smallest one that is still good enough.

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top