# Inkling 2 Mini — Thinking Machines' 26.4B looped open model for local use

> A small open MoE that loops its middle layers to reason better without paying for more stored weights.

- Org: Thinking Machines Lab
- Date: 2026-11-20
- Category: model
- Source: https://thinkingmachines.ai/news/inkling-2-mini/
- Canonical page: https://ai-tldr.dev/releases/thinking-machines-inkling-2-mini/

## In short

Inkling 2 Mini is a 26.4B-total, 3.4B-active open-weight MoE from Thinking Machines Lab. It runs its middle layers three times per token, fits in about 20GB and scores 79.2% on SWE-bench Verified.

## What it is

Inkling 2 Mini is the smallest model in Thinking Machines Lab's second Inkling generation, released November 20, 2026 with Apache 2.0 weights on Hugging Face. It has 26.4B total and 3.4B active parameters, takes text and image input and supports up to 1M tokens of context.

## How it works

The model stores 24 layers in four blocks of six, and the two middle blocks run three times per token, so 24 stored layers act like 48. Thinking Machines says looping the middle layers beat looping the last layers on 27 of 28 evaluations. Training started with 384 experts per layer; two-thirds were pruned to 128 and the model was healed with distillation on 1T tokens. Most weights ship in NVFP4, the precision they were trained in.

## Why it matters

Looping gives a small download more compute per token, which is aimed at people who run models on their own hardware. At about 20GB the model runs on a Mac with 32GB of RAM through Ollama, yet the lab reports 79.2% on SWE-bench Verified against 77.8% for Qwen3.8-27B. The lab also notes it trails larger models on HLE-Diamond and CritPt v2.

## Who it's for

developers running local coding and agent models

## Try it

vllm serve "thinkingmachines/Inkling-2-Mini"

## Quick facts

- Maker: Thinking Machines Lab
- Parameters: 26.4B total · 3.4B active
- Context window: Up to 1M tokens
- Input / output: Text and images in, text out
- License: Apache 2.0
- Availability: Hugging Face weights, Tinker, Ollama, GGUF

## FAQ

### Can Inkling 2 Mini run on a laptop?

Inkling 2 Mini needs about 20GB of memory, according to Thinking Machines Lab. The lab lists Ollama support on Macs with at least 32GB of RAM, GGUF builds through Unsloth and llama.cpp, and NVIDIA DGX Spark. For servers, the Hugging Face model card gives vLLM, SGLang and Docker Model Runner commands.

### How does Inkling 2 Mini compare with Qwen3.8-27B?

Thinking Machines Lab reports that Inkling 2 Mini scores 79.2% on SWE-bench Verified against 77.8% for Qwen3.8-27B. Qwen3.8-27B is ahead on Terminal Bench v2.1, 73.0% against 71.6%, and on ComplexConstraints v1.1, 35.6% against 33.4%. Inkling 2 Mini has only 3.4B active parameters per token.

### What is the variable-loop research model released with Inkling 2 Mini?

Thinking Machines Lab also published TML-d24-varloop-research on Hugging Face, a research model that can run a different number of loops. In the lab's tests its no-chain-of-thought time horizon rose from 16 seconds at one pass to 37 seconds at three passes, and its chain of thought got shorter with more loops while staying readable.

### Where can I fine-tune Inkling 2 Mini?

Inkling 2 Mini is available on Thinking Machines Lab's Tinker platform for inference and fine-tuning, with a playground. The lab also lists reinforcement-learning fine-tuning support in SkyRL, Miles and NVIDIA AutoModel, and inference support in SGLang, vLLM, NVIDIA Dynamo and TokenSpeed.

## Links

- [Announcement](https://thinkingmachines.ai/news/inkling-2-mini/)
- [Model Card](https://thinkingmachines.ai/model-card/inkling-2-mini/)
- [Hugging Face Weights](https://huggingface.co/thinkingmachines/Inkling-2-Mini)
- [Hugging Face Blog](https://huggingface.co/blog/inkling2mini)
- [Varloop Research Model](https://huggingface.co/thinkingmachines/TML-d24-varloop-research)
- [HN Discussion](https://news.ycombinator.com/item?id=50581260)
