Skip to content
NLEN
Illustration: OpenAI pipeline found 948 Linux root bugs in one run: what autonomous bug hunting means

Inkling 2 Mini: an open Thinking Machines model that runs on a Mac

By Ivo Donker · — text: Claude (AI)

What is Inkling 2 Mini? It is a small language model from Thinking Machines with open weights under the Apache 2.0 license, released on November 20. It has 26.4 billion parameters and fits in about 20 GB, so it runs on a Mac with 32 GB of memory or more. Its unusual feature is that it runs its middle layers three times per token, doing more computation without writing more text. It is aimed mainly at coding and agents.

We read Thinking Machines' announcement, the model card on Hugging Face and the Ollama model page. The news spread through a post by @thinkymachines on X.

What happened

Thinking Machines earlier released Inkling 2 Small, which the company says has 594 billion parameters. Inkling 2 Mini is the small version for your own hardware. The model card gives November 20, 2026 as its release date. It takes text and images, not audio, and has a context window of up to 1 million tokens.

The model is a mixture of experts. According to the blog, 3.4 billion parameters are active per token; because layers repeat, that adds up to 5.4 billion weight activations per token. Ollama lists this as "5.4B active". The model has 24 layers in four blocks of six. The two middle blocks run three times, the first and last once. Most weights were trained and ship natively in NVFP4, with the last four layers in MXFP8.

Figures and availability

Figures and availability
FeatureValue (according to the maker)
Parameters26.4 billion total, 3.4 billion active, 5.4 billion activations per token
Memoryabout 20 GB; a Mac with 32 GB or more, or an NVIDIA DGX Spark
LicenseApache 2.0
Contextup to 1 million tokens
SWE-bench Verified (bash only)79.2 percent
DeepSWE v1.148.2 percent
ScreenSpot Pro76.5 percent
Available throughHugging Face, Ollama (from 0.40.17), vLLM, SGLang, Tinker

Thinking Machines is open about weak spots too. On MultiChallenge (50.0 percent) and ComplexConstraints (33.4 percent) Inkling 2 Mini scores below Qwen3.8-27B. The model card says it knows fewer facts than Inkling 2 Small and is more likely to state wrong information with confidence. Repeating layers improves reasoning but stores no extra knowledge.

What this means

For anyone who wants to run agents locally, this is a serious candidate. It targets longer tasks with tools and is available in Ollama right away, with ready-made launchers for Claude Code, OpenCode and Hermes Agent, among others. That fits earlier work with small models, such as Nemotron 3.5 Bolt E4B, which we tested on a Mac, and agents like Hermes Agent.

The looping also matters for safety research. A model that computes more without writing it down is harder to follow through its reasoning, so Thinking Machines also releases a variant with an adjustable number of loops for researchers. After the long wait for the bigger model's weights, which we covered in our article on Inkling 2 Small, these weights were released straight away. For comparison on the closed side, see Qwen4-Max.

What we did not verify

We did not run the model ourselves. All scores come from Thinking Machines and were not measured independently. That Inkling 2 Small has 594 billion parameters comes from a post by @thinkymachines on X; we did not check it in a technical source.

Checked on November 21, 2026.

This article was written with AI (Claude) and checked against the sources below.

Sources