Artificial intelligence start-up Thinking Machines Lab on Friday, 20 November 2026, released Inkling 2 Mini, a compact AI model with openly downloadable weights that the company says is "small enough to run on your laptop." The announcement pitches the model mainly at agentic coding, where an AI system writes, tests and fixes software with little human help, and reports a 79.2% score on the widely used SWE-bench Verified coding test.
What is Inkling 2 Mini?
Inkling 2 Mini is a smaller sibling of Inkling 2 Small, the company's previous release. It has 26.4 billion parameters in total, but because it uses a "mixture-of-experts" design, only about 3.4 billion are active for each word-piece (token) it processes. Its memory footprint is about 20GB.
The model also uses an unusual "looped" design: it runs its middle layers three times for every token. Thinking Machines says this is meant to improve reasoning, especially for coding agents. Counting the loops, about 5.4 billion weights are activated per token.
Where can people use it?
The weights are published on Hugging Face, along with a research variant. According to the company, the model can be used through its own Tinker service for running, fine-tuning and chatting with models, and it runs on popular local tools including Ollama (on Macs with at least 32GB of memory), llama.cpp and Unsloth, as well as server software such as vLLM, SGLang and NVIDIA Dynamo. Thinking Machines says it can run offline on a MacBook Pro and on NVIDIA's DGX Spark desktop machine. The announcement does not name the licence for the weights or give API pricing.
How does it perform?
All figures below are Thinking Machines' own reported results and have not been independently checked:
- SWE-bench Verified: 79.2%
- SWE-bench Pro v2 (Hard): 77.9%
- Terminal Bench v2.1: 71.6%
- BrowseComp: 79.0%
- MCP Atlas (public): 82.7%
- DeepSWE v1.1: 48.2%, which the company says is the best among the models under 100 billion parameters it compared
The company compared the model against other mid-sized open models including Gemma 4 31B and Qwen3.8-27B. It did not claim to win everywhere: Qwen3.8-27B led on several tests, including Terminal Bench v2.1 at 73.0%, and Inkling 2 Mini scored lower on instruction-following tests such as MultiChallenge (50.0%) and ComplexConstraints v1.1 (33.4%). On the very difficult HLE-Diamond reasoning test it scored 15.2%.
Why does a laptop-sized model matter?
Most of the strongest AI models run only in large data centres, so users must send their data to a provider's servers and pay per use. A capable model that fits on a single high-end laptop lets developers, researchers and companies run it on their own hardware, including without an internet connection. Thinking Machines also highlighted energy efficiency, saying the model scored well on intelligence per watt and could run several persistent agents locally.
For India, where questions about data control and dependence on foreign AI providers are part of a live debate over what sovereign AI should mean, open-weight models that run locally are one practical route for developers who want to keep code and data on their own machines.
What next?
With the weights public, outside developers can now test the company's benchmark claims for themselves, and independent evaluations are likely to determine how the model is received. The 32GB memory requirement for Macs also means it remains out of reach for typical budget laptops.




