Skip to main content
v2026.11,986 entries · CC-BY 4.0

Editorial · CASRAI · AI and ML research outputs

Thinking Machines releases Inkling 2 Mini, an open-weights looped model for laptops

Thinking Machines Lab released Apache 2.0 open-weights Inkling 2 Mini on Nov 20, a looped model with about 20 GB of weights built to run on laptops.

Published 22 Nov 2026· 7 minute read
Close-up of a backlit laptop keyboard with a Dell laptop screen edge visible above it
Photo: Gre123gr / Wikimedia Commons, CC0 (cropped). Source · Licence

On November 20, 2026, Thinking Machines Lab released Inkling 2 Mini, a looped open-weights model it describes as small enough to run on a laptop, a new development after the lab’s release of the open-weights Inkling 2 Small on November 12. The weights are licensed under Apache 2.0, which matters to research teams that need to run a capable model on their own hardware.

What happened

Thinking Machines Lab published Inkling 2 Mini on November 20, 2026, with full weights on Hugging Face, according to the company’s announcement. The model is also offered through the company’s Tinker platform for inference and fine-tuning, and can be tried in the Tinker Playground. The Hugging Face model card lists the licence as Apache 2.0, with a separate Model Acceptable Use Policy published by the company.

Independent evaluator Artificial Analysis published its assessment the same day and described Inkling 2 Mini as an open-weights looped transformer released at an Intelligence Index score of 20. The Unite.AI report also dated November 20 repeats the company’s claim that the model is built for consumer laptops.

Key details

The company says Inkling 2 Mini is a looped transformer built as a Mixture-of-Experts model. It stores 24 layers in four blocks of six, and the two middle blocks run three times, which gives 48 layer applications per token without adding stored weights. According to the announcement, the model has 26.4 billion total parameters, 3.4 billion active per token, and about 5.4 billion weight activations per token once the repeated passes are counted. Artificial Analysis describes it as 26B total and 5.4B active, noting that the active figure includes three passes through the looped layers and that it uses 3.1B unique weights per text token.

Item Reported detail Source
Release date November 20, 2026 Thinking Machines Lab
Licence Apache 2.0 plus a separate acceptable use policy Hugging Face model card
Memory footprint About 20 GB; the model card recommends 32 GB or more of memory or VRAM Thinking Machines Lab; Hugging Face
Inputs and context Text and image input, 1 million token context Artificial Analysis; Hugging Face blog
Local options Ollama on Macs with at least 32 GB RAM; GGUF builds via Unsloth and llama.cpp Thinking Machines Lab
Precision Most weights NVFP4, last four layers MXFP8 Thinking Machines Lab

The announcement says most weights use NVFP4 and the last four layers use MXFP8, so the precision used in deployment matches the precision used in training. For a demonstration it reports running the model offline on a MacBook Pro with an M5 Pro chip and 48 GB of memory, and says Stanford tests used an NVIDIA DGX Spark with 128 GB.

Artificial Analysis reports that the model accepts text and image input, supports a 1 million token context and is available under Apache 2.0. The company’s own comparison table gives scores such as 79.2% on SWE-bench Verified and 48.2% on DeepSWE v1.1; those are company-reported figures, and the Unite.AI report notes that evaluation setups vary by model.

Background and context

Thinking Machines Lab released its first open-weights Inkling model in July 2026. On November 5 it introduced Inkling 2 Small, a 594-billion-parameter model that launched on Tinker with closed weights before the weights were posted under Apache 2.0, as covered in the earlier casrai.org report. Inkling 2 Mini is a much smaller model: the company says its training began with 384 experts per layer, two-thirds of which were pruned to leave 128, after which the model trained on 1 trillion tokens distilled from a larger internal Inkling model.

The looping design is the main technical difference. The company says looping the middle layers gave larger accuracy gains than looping the last layers on 27 of 28 evaluations. In the Unite.AI summary of the announcement, an unpruned non-looping model used 7.5% less training compute but needed 140% more inference compute to match the looped model’s loss on 512,000-token agentic traces. A research companion, TML-d24-varloop-research, lets users choose one or two loop passes instead of three, trading some quality for speed, as the Hugging Face blog explains.

Reaction and independent assessment

Artificial Analysis says Inkling 2 Mini matches Qwen3.8 27B (xhigh) on its Intelligence Index with about five times fewer active parameters, and sits on the Pareto frontier for intelligence against both total and active parameters. It reports an average of about 152,000 output tokens per task, roughly 38% more than Qwen3.8 27B, but a cost of $0.59 and 9.0 minutes per task against $3.53 and 38.2 minutes for that model, using Thinking Machines’ first-party pricing. It adds that its compute figure is a proxy, and that GPT-6 Luna (max) scores 28 at $0.15 per task.

The Hugging Face blog, written with the release, says the model leads in 32 of 48 benchmarks against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL, and that in informal tests its authors found it at least comparable to Qwen 3.8 27B while saying more validation is needed.

The company states its own limits. It says Inkling 2 Mini is not the strongest small model for multi-turn chat or instructions with many constraints, scoring 50.0% on MultiChallenge and 33.4% on ComplexConstraints v1.1, behind Qwen3.8-27B, and that looping adds no stored knowledge. The model card adds that it is less capable than Inkling 2 Small on factual recall, may state incorrect information confidently, sometimes gets stuck in long reasoning loops, and may take destructive actions such as deleting or overwriting files without confirmation.

What it means for researchers, universities and research offices

An Apache 2.0 licence and a roughly 20 GB checkpoint put a capable model within reach of a single workstation or a recent high-memory laptop, which gives research groups a route to running a model locally. The company’s recommendation of 32 GB or more of memory is the practical threshold to plan around. Where data cannot leave an institution, running the weights locally avoids sending prompts to an external service; research offices and data-protection teams still need to confirm that the acceptable use policy and their own rules cover the intended work. The open-weights model entry explains what a weights release does and does not include, and the guide to choosing and governing LLMs for research sets out the questions a research office should ask before adoption.

Reproducibility is the second consideration. A fixed set of downloadable weights can be pinned, archived and cited in a methods section, unlike a hosted model that can change without notice. The published model card is the document to record alongside it, including the number of loop passes if the variable-loop research checkpoint is used, since the company says that setting changes speed and may change quality.

The model card’s own warnings bear directly on research use: it recommends limiting output tokens and agent turns, running agentic workloads in a sandbox, requiring confirmation before destructive or irreversible actions, and validating the model for the domain before any medical, legal or safety-critical use. Unite.AI notes that looping can move reasoning out of written text, which complicates chain-of-thought monitoring. The company reports preliminary monitorability results on the variable-loop model: the length of reasoning without written chain of thought rose from 16 seconds at one pass to 37 seconds at three, and agentic sabotage rates did not trend upward with more loops.

What happens next

The weights, GGUF builds and inference support are available now, and the company lists support in SGLang, vLLM, Ollama, Unsloth, NVIDIA Dynamo, TokenSpeed and the Llama App, with fine-tuning through Tinker, SkyRL, Miles and NVIDIA AutoModel. The Hugging Face blog notes day-zero SGLang support and MLX support upstreamed to mlx-lm. Independent benchmarks beyond the first Artificial Analysis release have not been published at the time of release, so results from other evaluators will show how the company-reported scores hold up on other tasks. For safety-minded readers, the CAISI assessments of open-weight models show how government evaluators approach models with public weights.

Sources

Ask CASRAI · full access

Ask about this story

Full access: large monthly usage, up to 10 source passages behind each answer, and MCP and API access for Claude Code, Codex and other CLIs. Cancel any time.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →