Skip to content
AI News

Inkling 2 Mini: Thinking Machines Lab's Open Model That Runs on a Laptop

Presscart Team
• • Updated November 20, 2026

The short answer

Inkling 2 Mini is an open-weights AI model that Thinking Machines Lab released on Nov. 20, 2026. It has 26.4 billion parameters, takes text and images, and is about 20GB, small enough to run on a Mac with 32GB of memory. Its middle layers run three times for every token, a design called a looped transformer. Independent benchmarker Artificial Analysis found it matches Alibaba's Qwen3.8-27B on overall intelligence at about a sixth of the cost per task. The weights are free to download under an Apache 2.0 license.

AI search and open model news

What is Inkling 2 Mini?

Inkling 2 Mini is the smallest model in Thinking Machines Lab’s Inkling 2 family. It is a mixture-of-experts model: of its 26.4 billion total parameters, 3.4 billion are active for each token, according to the model card. It accepts text and image inputs, produces text, and supports a context of up to 1 million tokens, according to the vLLM and Hugging Face launch posts.

The lab built it for developers making agents, coding assistants, chatbots and retrieval systems. It is the companion to Inkling 2 Small, a much larger 594 billion parameter model the lab launched on Nov. 5 and released as open weights a week later. Mini borrows Small’s components but is designed for local machines, not data centers.

What is a looped transformer?

A looped transformer runs some of its layers more than once for each token, reusing the same weights. Inkling 2 Mini has 24 physical layers. Layers 6 to 17 run three times, so each token passes through 48 layers of computation while the model stores only 24, Hugging Face engineers explained in their launch post.

Each pass reads a gated message from the pass before, so every loop refines the previous one’s work. This lets the model reason more per token, according to the lab, and reach correct answers in fewer tokens, which helps most on long, multi-step tasks like agentic coding. The extra passes add about 1 billion effective active parameters each, for 5.4 billion per token in total.

Looping has a known limit. It improves reasoning but does not let the model store more facts, which depend mostly on total size. To make up for it, Mini leans on search and tools, and the lab reports it leads similar-sized open models on BrowseComp and DeepSearchQA.

How does Inkling 2 Mini compare with Qwen and Gemma?

Inkling 2 Mini matches Alibaba’s Qwen3.8-27B on overall intelligence while using about five times fewer active parameters, according to independent benchmarker Artificial Analysis. It scored 20 on the Artificial Analysis Intelligence Index, level with Qwen3.8-27B, and Artificial Analysis found it about six times cheaper and four times faster per task: $0.59 and 9.0 minutes, against $3.53 and 38.2 minutes for Qwen. On the lab’s own figures, Mini also leads comparable open models on coding and agentic search.

Model (lab-reported) Active / total params (B) DeepSWE v1.1 SWE-bench Verified BrowseComp MMMU Pro
Inkling 2 Mini 5.4 / 26.4 48.2% 79.2% 79.0%* 74.9%
Qwen3.8-27B 27.0 / 27.0 42.2% 77.8% 61.6% 76.3%
Nemotron 3.5 Super VL 12.0 / 124.5 24.7% 71.1% 78.9% 72.3%
Gemma 4 31B 30.7 / 30.7 0.3% 46.1% 37.0% 73.4%

*The model card notes Mini’s BrowseComp result was run with compaction, a context-management setup, so it may not match how the other models were tested.

Qwen3.8-27B scores higher on some knowledge-work and automation benchmarks, including GDPVal-AA and AutomationBench, and on the MMMU Pro vision test shown above. Mini also writes about 38% more output tokens than Qwen for the same score, Artificial Analysis found, though its smaller active size means it still uses about 2.8 times less compute.

How much does Inkling 2 Mini cost?

The weights are free to download and run on your own hardware. Through Thinking Machines Lab’s own API, Artificial Analysis lists pricing of $0.15 per million input tokens, $0.50 per million output tokens and $0.015 per million cached input tokens. At those rates, it measured an average cost of $0.59 per task on its Intelligence Index, against $3.53 for Qwen3.8-27B.

Can Inkling 2 Mini run on a laptop?

Yes. The checkpoint is about 20GB, stored mostly in the 4-bit NVFP4 format with the last four layers in MXFP8. The model card lists Apple silicon MacBooks, Mac minis and Mac Studios with 32GB or more of memory, a single Nvidia DGX Spark, and other GPUs with 32GB or more of video memory.

The lab’s researchers showed it working offline. On a flight, they ran it on a MacBook Pro in airplane mode and had it build a playable flight simulator in a few minutes.

Is Inkling 2 Mini open source, and where can you get it?

The weights are released under the Apache 2.0 license, which allows commercial use, and can be downloaded from Hugging Face. The lab also released a research version, TML-d24-varloop-research, that lets users choose the number of loops.

Support arrived on launch day across the main open-source tools. The model runs on vLLM, SGLang, Ollama, llama.cpp, Unsloth, NVIDIA Dynamo, TokenSpeed and the Llama App, and Hugging Face added transformers and MLX support. Developers can also fine-tune and run it through Tinker, Thinking Machines Lab’s training platform, or chat with it in the Tinker Playground.

Is a looped model harder to oversee?

Possibly, which is why the lab studied it. “Looping allows the model to reason longer without writing its reasoning down,” the lab wrote. That matters because one way to oversee a model is to read its written chain of thought.

Using the variable-loop research model, the lab reports that the written reasoning got shorter as loops were added but stayed readable. It also says the model did not score higher on tests of agentic sabotage as loops increased. The lab gave interpretability lab Goodfire early access to its looped models and is working with the National Deep Inference Fabric to make them available to researchers.

Presscart analysis: what the launch-day numbers actually say

The short version: Inkling 2 Mini is a real step down in the cost of running a capable model, but two of the launch claims need context. We read the lab’s announcement and model card against Artificial Analysis’ independent results, and checked how Google, ChatGPT and Gemini answered questions about the model a few hours after release.

“Fewer tokens” is measured against itself. The lab says looping helps the model reach the right answer with fewer tokens. Its own charts compare the research model at one to five loops, not Inkling 2 Mini against other models. Against Qwen3.8-27B, Artificial Analysis found the opposite on tokens: Mini writes about 38% more output tokens for the same score. Where it wins is compute, at about 2.8 times less, and price, at about a sixth of the cost per task. So the question isn’t tokens. It’s compute, the amount of intelligence you can run for a dollar, and that is the number agent builders should watch.

It is more likely to make things up when it doesn’t know, by the lab’s own numbers. AA-Omniscience measures how often a model avoids inventing an answer when it lacks the facts. In the model card, Mini’s non-hallucination rate on that test is 63.4%, against 71.4% for Qwen3.8-27B. The card also lists degraded performance in long multi-turn conversations. Run it with search and tools, and check what it tells you. A model is a tool. It’s not an expert.

How did AI search engines answer questions about Inkling 2 Mini on launch day?

All three answered “What is Inkling 2 Mini?” correctly on the core facts within hours: 26.4 billion parameters, 3.4 billion active, a looped design, about 20GB, Apache 2.0. What they cited tells you who gets credit for news in AI search.

Engine (Nov. 20, about 8 p.m. ET) Sources cited
ChatGPT Thinking Machines Lab, Hugging Face, Nvidia Dynamo docs, Ollama
Gemini Artificial Analysis, Hugging Face, Thinking Machines Lab, Unsloth
Google AI Overview Reddit, The Defiant, Unite.AI, OrcaRouter

None of the three cited a major news outlet. The first pages to publish clear, specific facts got the credit: the lab, the model hosts, the tool makers and one independent benchmarker. Google’s AI Overview also said a separate acceptable-use policy restricts uses such as employment and healthcare decisions. We could not find that policy on the model card or in the announcement, so treat it as unconfirmed.

Method: one answer per engine for one prompt, US English, collected through DataForSEO on Nov. 20, 2026. AI answers change from run to run, so this is a snapshot, not a ranking.

Who is Thinking Machines Lab?

Thinking Machines Lab is an AI company founded by Mira Murati, its chief executive. It released Tinker, its fine-tuning platform, in October 2025, and shipped Inkling 2 Small two weeks before Mini. Business Insider reported on Nov. 6, citing people familiar with the matter, that the company was raising $5 billion at a $40 billion pre-money valuation in a round led by Accel. The company had not announced the round at publication.

Sources

  1. Thinking Machines Lab, “Introducing Inkling 2 Mini,” Nov. 20, 2026: https://thinkingmachines.ai/news/inkling-2-mini/
  2. Inkling 2 Mini model card, Hugging Face: https://huggingface.co/thinkingmachines/Inkling-2-Mini
  3. Hugging Face, “Welcome Inkling 2 Mini,” Nov. 20, 2026: https://huggingface.co/blog/inkling2mini
  4. vLLM, “Inkling 2 Mini on vLLM: Day-0 Support for a Looped Model,” Nov. 20, 2026: https://vllm.ai/blog/2026-11-20-inkling2-mini
  5. Thinking Machines Lab, “Introducing Inkling 2 Small,” Nov. 5, 2026: https://thinkingmachines.ai/news/inkling-2-small/
  6. Artificial Analysis, “Thinking Machines has released Inkling 2 Mini, an open weights looped transformer,” Nov. 20, 2026: https://artificialanalysis.ai/articles/inkling-2-mini
  7. Thinking Machines Lab, “Announcing Tinker,” Oct. 1, 2025: https://thinkingmachines.ai/news/announcing-tinker/
  8. DataForSEO captures of Google AI Overview, ChatGPT and Gemini answers, Nov. 20, 2026 (raw responses kept on file)
  9. Business Insider, Nov. 6, 2026: https://www.businessinsider.com/thinking-machines-lab-is-closing-5-billion-new-round-2026-11

Frequently Asked Questions

The what, why, and how of Presscart.

When was Inkling 2 Mini released?

Thinking Machines Lab released Inkling 2 Mini on Nov. 20, 2026.

How many parameters does Inkling 2 Mini have?

It has 26.4 billion total parameters, with 3.4 billion active per token, or 5.4 billion effective active parameters when its three loops are counted.

What hardware do you need to run Inkling 2 Mini?

About 32GB of memory. The model card lists Apple silicon Macs with 32GB or more, a single Nvidia DGX Spark, and GPUs with 32GB or more of video memory.

Is Inkling 2 Mini free to use commercially?

Yes. It is released under the Apache 2.0 license.

Is Inkling 2 Mini better than Qwen3.8-27B?

They score the same, 20, on the Artificial Analysis Intelligence Index, but Inkling 2 Mini was about six times cheaper and four times faster per task. On the lab's own benchmarks it leads on most coding and agentic search tests, while Qwen3.8-27B leads on some knowledge-work and vision tests.

How much does Inkling 2 Mini cost?

The weights are free. Through Thinking Machines Lab's API, Artificial Analysis lists $0.15 per million input tokens and $0.50 per million output tokens.

Does Inkling 2 Mini hallucinate more than Qwen3.8-27B?

On AA-Omniscience, which measures how often a model avoids inventing answers it doesn't know, the model card reports a 63.4% non-hallucination rate for Inkling 2 Mini, against 71.4% for Qwen3.8-27B. It does better with search and tools.

PT

Presscart Team

Editorial Team

The Presscart editorial team covers PR strategy, media placements, and the tools founders and marketers use to get press coverage.

Tags: #ai news #open weights #thinking machines lab #local llm #ai search
Share:

More news

AI News

AI and market research news

Can AI Replace Market Research Surveys? What the Colgate and PyMC Labs Synthetic Consumer Study Shows

Presscart Team