Skip to content
Neural Daily

ModelsNews

Inkling 2 Mini: open 26B model that runs on a laptop

🇭🇺 Olvasd el magyarul →

Thinking Machines released Inkling 2 Mini, a 26.4B-parameter Apache 2.0 MoE model that runs locally in about 20 GB of memory. It is a looped transformer that runs its middle layers three times per token.

· 3 min read

A Thinking Machines Lab fekete logója fehér háttéren
Image: Thinking Machines Lab · Press kit · Source

What was released?

Thinking Machines, the lab founded by Mira Murati, released Inkling 2 Mini on November 20 (Thinking Machines). It is the little sibling of the recently launched Inkling 2 Small (Inkling 2 Small: a 594B MoE from Thinking Machines), sized for a laptop. The weights are on Hugging Face under Apache 2.0 (Hugging Face).

  • Mixture-of-Experts: 26.4B total parameters, 3.4B active.
  • Text and image input, text output; aimed at agentic, tool-use and coding work.
  • Most weights are trained and deployed natively in NVFP4, the last four layers in MXFP8. Total memory footprint is about 20 GB.
  • The model card recommends Apple silicon with 32 GB or more, a DGX Spark, or a GPU with 32 GB+ of VRAM.

What is a looped transformer?

Inkling 2 Mini runs its middle layers three times for every token, giving 5.4B weight activations per token. The model can reason longer without writing its reasoning down. The lab says this helps most in agentic coding, where reasoning quality matters more than stored knowledge.

Looping does not add factual knowledge: its AA-Omniscience score is in line with models of the same size. Thinking Machines also released a research variant with a variable loop count. Its no-CoT time horizon rose from 16 seconds at one pass to 37 seconds at three, its chain of thought got shorter but stayed legible, and it did not score higher on agentic sabotage tests.

How good is it?

By the lab's own numbers, Inkling 2 Mini leads all open-weights models under 100B total parameters on DeepSWE v1.1 (48.2%) and Vals Index v2.2 (46.5%). It is also ahead of similarly sized models on SWE-bench Pro v2 and trails only Qwen3.8-27B on some other tests. On PMPP-Hard, a kernel-optimization benchmark, it approaches the much larger Inkling 2 Small.

In the demo, researchers ran it offline on a 48 GB M5 Pro MacBook Pro during a flight and had it build a playable flight simulator in a few minutes.

Where does it run?

Locally on SGLang, vLLM, Ollama, llama.cpp (GGUF), Unsloth, NVIDIA Dynamo, TokenSpeed and Llama App. RL fine-tuning is supported in SkyRL, Miles and NVIDIA AutoModel, and the hosted route is the lab's Tinker platform for inference and fine-tuning.

Why it matters

This is one of the first widely available models with a looped architecture. If you want a local coding agent with no cloud, or a small model to fine-tune on company data, it is now one of the strongest open options under 30B. Test it against Qwen3.8-27B on your own tasks, since the published scores come from the lab's own runs with varying setups (Harness adapters: RL-tuned LoRAs lift Qwen3.8-27B agents).

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles