# Thinking Machines releases Inkling 2 Mini, a 26B model that loops

> The Apache 2.0 model runs its middle layers three times per token, fits in about 20GB and ties Qwen3.8-27B on the Artificial Analysis index.

URL: https://claudegpt.si/news/inkling-2-mini-open-weights  
Published: 2026-11-20T22:51:15.138Z  
Updated: 2026-11-20T23:00:33.165Z  
Author: ClaudeGPT Si Newsroom  
Section: AI Models  
Topics: Thinking Machines Lab, Inkling, Open-weight models, Language models  
Publisher: ClaudeGPT Si

## Key takeaways

- Thinking Machines Lab released Inkling 2 Mini on November 20 as open weights under Apache 2.0.
- It has 26.4B parameters, 3.4B active per token, and runs its middle 12 layers three times for each token.
- Artificial Analysis scores it 20.1 on its Intelligence Index, tied with Qwen3.8-27B.

Thinking Machines Lab released Inkling 2 Mini on November 20, an open-weight model with 26.4 billion parameters that fits in about 20GB and runs on a laptop with 32GB of memory. The [announcement](https://thinkingmachines.ai/news/inkling-2-mini/) says it is a looped transformer: its middle layers run three times for every token, which the lab says buys reasoning without making the model bigger. The weights are on [Hugging Face](https://huggingface.co/thinkingmachines/Inkling-2-Mini) under the Apache 2.0 licence.

The interesting part is the trade. Most small models get better by adding parameters or by thinking out loud for longer. Inkling 2 Mini reuses the same layers several times per token instead. If the benchmark numbers hold up, that is a cheap way to get more out of a model that has to fit in consumer memory.

## How the looping works

Inkling 2 Mini is a mixture-of-experts model, which means each token goes to only a few of many small sub-networks. Each token uses 8 of 128 experts, so only 3.4 billion of the 26.4 billion parameters are active at a time.

The model stores 24 layers in four blocks of six. The first and last blocks run once. The two middle blocks run three times, so each token passes through 48 layers. Counting those repeats, about 5.4 billion weights are used per token.

Thinking Machines says it tested where to loop. Looping the middle 12 layers beat looping the last 12 on 27 of 28 evaluations. It also says a version without loops was 7.5% cheaper to train but needed 140% more compute at inference to reach the same loss on long agent transcripts.

The weights ship in NVFP4, a 4-bit number format, for 20 layers and MXFP8 for the last four. The model was trained in those formats, so there is no extra quality loss from shrinking it afterwards, according to the lab. The [model card](https://huggingface.co/thinkingmachines/Inkling-2-Mini) lists a context window of up to 1 million tokens.

## Benchmarks put it level with Qwen3.8-27B

The lab compares Inkling 2 Mini with Qwen3.8-27B, Gemma 4 31B, Muse Glimmer and Nemotron 3.5 Super VL. Its own table shows a split result:

| Benchmark | Inkling 2 Mini | Qwen3.8-27B |
| --- | --- | --- |
| SWE-bench Verified (bash only) | 79.2% | 77.8% |
| SWE-bench Pro v2 (Hard) | 77.9% | 73.8% |
| DeepSWE v1.1 | 48.2% | 42.2% |
| Terminal Bench v2.1 | 71.6% | 73.0% |
| MMMU Pro | 74.9% | 76.3% |
| HLE-Diamond | 15.2% | 15.3% |

The independent number is the Artificial Analysis Intelligence Index, where Inkling 2 Mini scores 20.1, tied with Qwen3.8-27B. Artificial Analysis noted that it gets there with about five times fewer active parameters. That matters for speed and cost per token, since active parameters decide how much work each token takes.

Thinking Machines is frank about the weak spots. "Inkling 2 Mini is not the strongest small model for multi-turn chat or for instructions with many constraints," the post says. It scores 50.0% on MultiChallenge, behind Qwen3.8-27B, according to the post. Extra loops also do not help on knowledge tests, which depend on total size.

## Where to run it

The weights work with Transformers, vLLM, SGLang, llama.cpp, Ollama and Unsloth, which already posted GGUF builds. Ollama on a Mac needs at least 32GB of memory. Thinking Machines also serves the model on Tinker, its own platform for inference and fine-tuning. The post gives no per-token API price.

The lab also released a research variant, TML-d24-varloop-research, trained to run between one and five passes, so researchers can study what more loops do to reasoning and to how readable the model's chain of thought stays.

This is the second Inkling 2 release this month, after [Inkling 2 Small](https://claudegpt.si/news/inkling-2-small-thinking-machines). The next thing to watch is whether independent coding tests, outside the lab's own harness, match the SWE-bench numbers it reports.

Correction, November 20: An earlier version of this article said Inkling 2 Mini trails Qwen3.8-27B on ComplexConstraints. The announcement's text says so, but its own table shows Inkling 2 Mini ahead, 33.4% to 30.1%, so that comparison has been removed.

## Sources

- [Inkling 2 Mini](https://thinkingmachines.ai/news/inkling-2-mini/), Thinking Machines Lab
- [thinkingmachines/Inkling-2-Mini model card](https://huggingface.co/thinkingmachines/Inkling-2-Mini), Hugging Face
- [Artificial Analysis on Inkling 2 Mini (via Techmeme)](https://www.techmeme.com/261120/p37#a261120p37), Techmeme
