Thinking Machines Lab released Inkling 2 Mini on 20 November 2026, and it breaks the usual trade-off for small open models: it is a 26.4-billion-parameter model with open weights under Apache 2.0, it fits in about 20GB, it runs on a Mac with 32GB of memory — and on the benchmarks that matter for agentic coding, it leads every open-weights model with under 100 billion total parameters. The trick is looping: the middle 12 of its 24 layers run three times for every token, so the model does more reasoning per token without carrying more weights.
Every benchmark in this article is the vendor’s own, reported by Thinking Machines Lab or measured by Artificial Analysis from the lab’s first-party pricing. No independent reproduction had landed by 22 November. Treat the scores as the lab’s claim, not a settled verdict — but the claim is specific, the numbers are checkable, and the model is downloadable today, so anyone can test it.
What is Inkling 2 Mini?
Inkling 2 Mini is a mixture-of-experts model from Thinking Machines Lab, the AI startup founded by former OpenAI chief technology officer Mira Murati. It takes text and images as input, replies in text, and carries a 1-million-token context window. The full weights are available on Hugging Face under the Apache 2.0 licence, alongside a separate Model Acceptable Use Policy published by the company. It is also available for inference and fine-tuning on the lab’s Tinker platform, and for chat in the Tinker Playground.
| Specification | Inkling 2 Mini |
|---|---|
| Total parameters | 26.4 billion |
| Active parameters per token | 3.4 billion (5.4 billion effective, counting the loops) |
| Architecture | Looped transformer, mixture of experts — 24 layers, middle 12 run three times |
| Memory footprint | About 20GB (NVFP4 for 20 layers, MXFP8 for the last four) |
| Context window | 1 million tokens |
| Input / output | Text and images in, text out |
| Licence | Apache 2.0, plus a separate acceptable use policy |
| Hosted API price | $0.15 input / $0.50 output per 1 million tokens |
| Runs on | Mac with 32GB+ RAM via Ollama; NVIDIA DGX Spark; GGUF builds via Unsloth and llama.cpp |
The size is the headline. The lab’s bigger model, Inkling 2 Small, needs hundreds of gigabytes of VRAM for its NVFP4 checkpoints and is still closed-weight while external safety testing runs. Mini is the version you can actually download and run yourself — on hardware many developers already own.
How the looping works — and what it buys you
A normal transformer passes each token through its layers once, top to bottom. Inkling 2 Mini has 24 layers divided into four blocks of six. The first and last blocks run once; the two middle blocks run three times. That gives 48 layer applications per token from 24 stored layers — the model re-reads its own middle layers, and each pass also reads the previous pass’s output, which the lab calls the message.
The design rests on an old intuition, borrowed from the 2018 Universal Transformer paper: the middle layers carry the abstract reasoning, while the first and last layers mostly translate between words and internal representations. Looping the middle layers, the lab found, was worth far more than looping the last ones. At training step 2,000, the no-loop baseline would have needed 1.51–1.91 times as much compute to match the middle loop’s loss, against 1.24–1.36 times for the last loop. And a non-looping version of the same model cost 7.5% less to train but needed 140% more inference compute to match the looped version’s loss on 512,000-token agentic traces.
Two things looping does not do. It does not add knowledge: the model stores facts in its weights, and 26.4 billion parameters hold about as many facts as any model of that size — its score on AA-Omniscience, a factuality measure, is in line with other models of the same size, and closed-book factual questions barely improve with more loop passes. And it does not make the model bigger: the weights are the weights. What looping buys is reasoning depth per token — more serial computation before the model has to write anything down.
That last point is why the lab released it at all. A model that reasons internally, without spelling out its chain of thought, is harder to oversee by reading that chain of thought — so Thinking Machines also released a variable-loop research model, TML-d24-varloop-research, trained with one to five loop passes, so researchers can study exactly how looping affects monitorability. Its early finding is reassuring at this scale: the chain of thought gets shorter as loops increase, but it stays legible. On GPQA Diamond questions answered correctly at all five pass counts, reasoning at five passes was less than half as long as at one pass (1,661 tokens against 3,703), with sentence length and dictionary-word share unchanged — the model is more direct and backtracks less. And the no-chain-of-thought time horizon — how long a task the model can complete at 50% reliability without writing any reasoning — rose from 16 seconds at one pass to 37 seconds at three.
How it scores against the rivals
Here is the part to read with the honesty line in mind: these are Thinking Machines Lab’s own numbers, from its comparison table published with the release. The rivals are the open-weights models of similar size — Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL.
| Benchmark | Inkling 2 Mini | Best rival | What it measures |
|---|---|---|---|
| DeepSWE v1.1 | 48.2% | 42.2% (Qwen3.8-27B) | Agentic coding, real-world software tasks |
| SWE-bench Verified | 79.2% | 77.8% (Qwen3.8-27B) | Fixing real GitHub issues |
| SWE-bench Pro v2 (Hard) | 77.9% | 73.8% (Qwen3.8-27B) | Harder software-engineering tasks |
| Vals Index v2.2 | 46.5% | 29.8% (Nemotron 3.5 Super VL) | Agentic evaluation suite |
| Artificial Analysis Intelligence Index v5.0 | 20.1 | 20.1 (Qwen3.8-27B, tied) | Overall intelligence, third-party index |
| BrowseComp | 79.0% | 78.9% (Nemotron 3.5 Super VL) | Deep web research with tools |
| MCP Atlas | 82.7% | 77.3% (Qwen3.8-27B) | Tool use across MCP servers |
Two numbers outside the lab’s own table are worth noting. Artificial Analysis, which runs its own Intelligence Index, puts Mini at 20 — matching Qwen3.8-27B and Nanbeige4.3-33B — while using about five and six times fewer active parameters respectively. And on cost per task, its measurement is stark: $0.59 and 9.0 minutes per Intelligence Index task for Mini, against $3.53 and 38.2 minutes for Qwen3.8-27B — roughly six times cheaper and four times faster, using about 2.8 times less compute, though it spends about 38% more output tokens to get there. Vals AI also reported that Mini ties Claude Opus 5.5 (CVP) on CyberBench v1.1 at 74.58% — a frontier closed model, matched on a cybersecurity benchmark, at a fraction of the cost per test.
Stanford researchers got an early look at a different angle — intelligence per watt. On an NVIDIA DGX Spark, Mini scored 0.703 on their energy-efficiency measure, against 0.430 for Qwen3.6-35B-A3B, 0.159 for Gemma-4-26B-A4B and 0.036 for Qwen3.5-27B. That is the number that matters if you want to run several persistent agents on local hardware: the lab says all 20 residents of its demo pixel town, Thinkyville, are managed by Mini running on a single offline MacBook.
Where it falls short
The honest version of this model includes its weak spots, and the lab publishes them in the same post. Mini is not the strongest small model for multi-turn chat or for instructions with many constraints: it scores 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, behind Qwen3.8-27B on both. It also trails larger models on reasoning-heavy benchmarks such as CritPt v2 and HLE-Diamond. And as noted above, looping does not help it remember more facts — if your task is closed-book knowledge, a bigger model still wins.
There is a practical caveat too. Support for looped models in local runtimes is new, and the model card itself warns that speed varies by runtime and hardware. The Ollama path needs version 0.40.17, which was still in pre-release on 22 November — pin the version in your scripts and expect checksum or manifest changes before the stable release.
How to run it on your laptop
The shortest path is Ollama on a Mac. You need an Apple silicon Mac with at least 32GB of unified memory and about 20GB of disk space for the download, plus Ollama 0.40.17 or later:
ollama pull inkling-2-miniollama run inkling-2-mini
That is the whole install. The lab’s own researchers ran it offline on a MacBook Pro (M5 Pro, 48GB) on a flight, where it built a playable flight simulator in a few minutes — the demo is on the announcement page. If you do not have a Mac, the other supported paths are a single NVIDIA DGX Spark, any GPU with 32GB or more of VRAM, or quantized GGUF builds through Unsloth and llama.cpp. Open-source inference is available on SGLang, vLLM, NVIDIA Dynamo, TokenSpeed and the Llama App, with MLX support upstreamed to mlx-lm. For fine-tuning, the lab lists Tinker, SkyRL, Miles and NVIDIA AutoModel.
If you would rather not run it locally, the hosted route is Tinker: inference and fine-tuning on the platform, chat in the Tinker Playground, and a first-party API at $0.15 per million input tokens and $0.50 per million output tokens. That is the price Artificial Analysis used for its $0.59-per-task figure — so the hosted cost and the local cost are the same story told two ways: this is a capable coding model that costs very little to run.
Who should try it — and who should skip it
Three numbers decide it: a 20GB footprint, $0.59 per benchmark task, and 48.2% on DeepSWE v1.1. If those three numbers describe your bottleneck — you want a coding agent that runs on your own machine, offline, without a cloud bill — Mini is the most interesting open model released this month. Try it first for agentic coding, terminal work and tool-heavy research, where its scores are strongest and where its looping actually pays.
| Try Inkling 2 Mini if… | Skip it (for now) if… |
|---|---|
| You want a coding agent that runs offline on a 32GB Mac or a DGX Spark | Your work is long multi-turn chat or heavily constrained instructions — Qwen3.8-27B still leads there |
| You need tool use and web research — BrowseComp 79.0%, MCP Atlas 82.7% | You need closed-book factual knowledge — looping adds reasoning, not facts |
| You want open weights you can fine-tune, under Apache 2.0 | You need a stable, version-pinned local runtime today — looped-model support is still new |
| You are a researcher studying looped transformers and chain-of-thought monitoring | You want the biggest model in the family — Inkling 2 Small is still closed-weight |
Our take: the looping idea is the real story here, not just another leaderboard entry. Small models have always been cheap and weak; Mini is cheap and, on coding, genuinely strong — because it spends its compute re-reading its own middle layers instead of writing longer chains of thought. The open question is the one the lab itself flags: as looped models get bigger, will their reasoning stay readable? Mini ships with the research model to answer that question. Download it, run the flight-simulator test on your own machine, and decide from your own numbers — the weights are there for exactly that.
Frequently Asked Questions
Can Inkling 2 Mini really run on a laptop?
Yes — on an Apple silicon Mac with at least 32GB of unified memory, via Ollama, with about 20GB of disk space for the checkpoint. The lab’s researchers ran it offline on a MacBook Pro with 48GB. Other supported hardware is a single NVIDIA DGX Spark or any GPU with 32GB or more of VRAM.
Is Inkling 2 Mini fully open source?
The weights are released under the Apache 2.0 licence, which permits commercial use, and the full weights are on Hugging Face. But the release also carries a separate Model Acceptable Use Policy from Thinking Machines Lab — read that before commercial use. The bigger sibling, Inkling 2 Small, is still closed-weight while external safety testing runs.
How is it different from Qwen3.8-27B?
Both are open-weights models of similar size, and they tie at 20 on the Artificial Analysis Intelligence Index. Mini leads on agentic coding (DeepSWE v1.1: 48.2% against 42.2%) and costs far less per task ($0.59 against $3.53 on Artificial Analysis’ measure), while Qwen3.8-27B leads on multi-turn chat and constraint-heavy instructions. The difference is architectural: Mini loops its middle layers three times per token; Qwen3.8 is a conventional transformer.
What hardware do I need to run it?
A Mac with 32GB or more of unified memory, a single NVIDIA DGX Spark, or a GPU with 32GB or more of VRAM. The checkpoint is about 20GB. Ollama 0.40.17 or later is the easiest Mac path; GGUF builds via Unsloth and llama.cpp cover other setups.
Can I use Inkling 2 Mini for commercial projects?
The Apache 2.0 licence permits commercial use, but the release also includes a separate acceptable use policy from Thinking Machines Lab that adds conditions. Check the policy on the Hugging Face model card before using the model in a commercial product.
Sources and methodology
Claims in this article rest on Thinking Machines Lab’s announcement “Introducing Inkling 2 Mini” (20 November 2026, read in full on 22 November 2026), Artificial Analysis’ “Thinking Machines has released Inkling 2 Mini” (20 November 2026), Unite.AI’s release report (20 November 2026), and explainx.ai’s Inkling 2 Mini write-up (updated 22 November 2026) for the Vals CyberBench result and the Ollama 0.40.17 pre-release note. All benchmark figures are vendor-reported — Thinking Machines Lab’s own table, or Artificial Analysis’ measurement using the lab’s first-party pricing — and no independent reproduction had been published by 22 November. This article is desk-researched, not hands-on tested: the model was not run locally for this piece. Ollama version and hardware requirements were cross-checked against the announcement and the model card on 22 November 2026.
Sources: Thinking Machines Lab — Introducing Inkling 2 Mini · Artificial Analysis — Thinking Machines has released Inkling 2 Mini · Unite.AI — Thinking Machines Lab releases Inkling 2 Mini for local devices · explainx.ai — Inkling 2 Mini: CyberBench 74.58%, Run on a Laptop
Related reading on OpenAIMaster: how to track what each coding-agent PR really costs with LangSmith and LangChain’s $300M raise and the new LangSmith Agent Platform.
Arva Rangwala covers AI news, AI tools, guides and prompts for OpenAIMaster — what is new, what is worth using, and how to put AI to work.


