Mira Murati's Thinking Machines has released Inkling 2 Mini, a free 26.4B-parameter open-weight model under Apache 2.0 that needs only about 20GB of memory.

Thinking Machines Lab, the San Francisco start-up founded by former OpenAI chief technology officer Mira Murati, released Inkling 2 Mini on 20 November. It is an open-weight model, meaning anyone can download the files and run them on their own hardware, and it is published on Hugging Face under the permissive Apache 2.0 licence. The model has 26.4 billion parameters in total but uses only 3.4 billion for each token, or small piece of text, thanks to a mixture-of-experts design. Its unusual feature is looping: of its 24 layers, the 12 in the middle run three times for every token, giving the model more reasoning depth without adding more weights. The lab says the whole model fits in about 20GB of memory.

Thinking Machines says Inkling 2 Mini leads all open-weight models under 100 billion parameters on several agentic coding tests, with 48.2% on DeepSWE v1.1 and 79.2% on SWE-bench Verified, against 77.8% for Alibaba's Qwen3.8-27B on the latter. Independent tester Artificial Analysis gives it 20 points on its Intelligence Index, which VentureBeat reports is level with Qwen3.8-27B, but finds it completes each test task for about $0.59, against $3.53 for the Qwen model. The lab is open about weak spots: the model trails Qwen on multi-turn chat and complex instructions, and its factual recall is limited by its size. Artificial Analysis calls it very verbose, and VentureBeat reports clipped phrasing in some non-English languages. The release follows the much larger, 594-billion-parameter Inkling 2 Small, launched on 5 November.

Developers can download the weights for free from Hugging Face and run them with popular local tools such as Ollama, vLLM and llama.cpp. According to reports, a Mac with at least 32GB of memory is enough, while community GGUF versions from Unsloth target machines with less. For those who prefer the cloud, Thinking Machines offers the model through its own API at $0.15 per million input tokens and $0.50 per million output tokens, according to Artificial Analysis and VentureBeat, and it can be customised on the company's Tinker fine-tuning platform. No official list of supported languages has been published, so Arabic-speaking users in Morocco and the wider region should test it on their own tasks first. The Apache 2.0 licence allows commercial use, alongside an acceptable use policy.