On November 20, 2026, Thinking Machines Lab released Inkling 2 Mini, the small version of the open model it put out earlier this month. The license is Apache 2.0, so you can use it for work and inside your own products. One difference from its bigger sibling matters most: Inkling 2 Mini is built to run on a desktop computer or a Mac, with no rented servers.
The model file is about 20 GB. According to the company, a Mac with Apple Silicon and at least 32 GB of memory is enough.
What is Inkling 2 Mini?
It's a language model that reads text and images and answers in text. Like Inkling 2 Small, it uses the MoE technique (Mixture of Experts): each layer holds 128 "experts", but only 8 of them work on each word. The technical news is something else. Of the model's 24 layers, the 12 in the middle run three times with the same weights. That way the model "thinks" more about each word without getting bigger to download. Thinking Machines Lab calls it a looped transformer, and it's the first one the company has released.
The key figures from the official model card:
| Feature | Inkling 2 Mini |
|---|---|
| Parameters | 26.4 billion total, 3.4 billion active (5.4 billion "effective" counting the three passes) |
| Input and output | Text and images in, text out |
| Context | Up to 1 million tokens |
| License | Apache 2.0 |
| Size | About 20 GB (NVFP4 and MXFP8 formats) |
| Listed hardware | Mac with Apple Silicon and at least 32 GB of memory, NVIDIA DGX Spark, other graphics cards with at least 32 GB of VRAM recommended |
| Release date | November 20, 2026 |
How does it compare with other small models?
Thinking Machines Lab compares it with Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL. According to the Hugging Face blog, which worked on the launch with the company, Inkling 2 Mini comes first in 32 tests out of 48. The closest rival is Qwen3.8-27B. Here are some figures claimed by the company, not independently verified:
| Test | Inkling 2 Mini | Qwen3.8-27B |
|---|---|---|
| SWE-bench Verified (coding) | 79.2% | 77.8% |
| BrowseComp (web research) | 79.0% | 61.6% |
| MCP Atlas (tool use) | 82.7% | 77.3% |
| OmniDocBench v1.6 (reading documents) | 92.8% | 91.5% |
| Terminal Bench v2.1 (terminal work) | 71.6% | 73.0% |
| AutomationBench-AA (automations) | 41.8% | 48.2% |
The picture is clear: it does better at research and tool use, it's level on code and documents, and it loses some ground on automations.
Where do you download it, and with which apps?
- Hugging Face: the official weights, plus 4-bit and 8-bit GGUF versions for llama.cpp.
- Ollama: the model is in the library as inkling-2-mini (20 GB); the official model card lists it for Macs with Apple Silicon.
- Unsloth and llama.cpp, with compressed versions, if you use a PC with a graphics card.
- For developers: Transformers, MLX, vLLM, SGLang and the company's Tinker platform.
If you've never used a model locally, start with our guide on how to install Ollama and try your first model.
What changes for people at work?
Inkling 2 Small, whose weights came out on November 12, has 594 billion parameters: to use it you need servers or a paid service. Inkling 2 Mini fits on a 32 GB Mac. That means keeping documents, contracts and client data in-house, without sending them to an outside service. If you handle confidential information, this is the point that matters most.
The company lists the limits itself, and they're worth knowing before you try it:
- it gets facts and general-knowledge questions wrong more often, sometimes with confidence;
- outside English, quality can drop;
- it sometimes gets stuck in long, repetitive reasoning, especially when fixing code;
- when it uses tools, it may delete or overwrite files without asking for confirmation.
Our take: for summaries, reading PDFs and searching your own documents, it's worth a try if you have a 32 GB Mac. We wouldn't trust its answers "from memory" without checking them, and we wouldn't give it free access to work folders. For background on running AI on your own machine, see our article on Inkling 2 Small.
Frequently asked questions
Does Inkling 2 Mini work on a Windows PC?
The official model card says other graphics cards are possible and recommends at least 32 GB of VRAM, which means high-end cards. For the compressed GGUF versions from Unsloth and llama.cpp, the company doesn't give a minimum memory.
Is it free?
Yes, the weights are a free download and the Apache 2.0 license also allows commercial use. You still have to follow the company's acceptable use policy.
How is it different from Inkling 2 Small?
Small is much bigger and, according to the company, does better in most tests, with the widest gap on facts and knowledge. Mini gives up some quality to run on a single computer.
Official sources
- Thinking Machines Lab: Inkling 2 Mini model card (November 20, 2026)
- Hugging Face: Welcome Inkling 2 Mini (November 20, 2026)
- Ollama: inkling-2-mini
Transparency: this article is written with the help of artificial intelligence models, starting from the official sources listed, and is reviewed before publication.
















