Thinking Machines Lab has published Inkling 2 Mini, an open-weight model designed to run on a personal computer. According to figures released by the lab, the model outperforms Gemini 2.5 Pro, Google's flagship from a year earlier, on coding tests under the conditions the company declares.
The model is a Mixture-of-Experts with 26.4 billion total parameters, 3.4 billion active per token. Its distinguishing feature is a looped structure: the middle layers run three times for every token. Thinking Machines says this lets the model reason more deeply without writing out each step, reaching the right answer with fewer tokens. The lab also argues this makes its internal reasoning harder to read, which matters for anyone monitoring models through their chain of thought.
The declared results
In the table published by the company, Inkling 2 Mini scores 48.2% on DeepSWE v1.1 and 46.5% on Vals Index v2.2, the highest among open models under 100 billion total parameters. On SWE-bench Verified, which measures fixes to real GitHub issues, it reaches 79.2% in bash-only mode. On Artificial Analysis, an independent index aggregating several tests, it scores 20.1 points, tied with Qwen3.8-27B. The comparison figures for Gemini 2.5 Pro come from a Reddit table, not from Google's own materials, which in its 2025 announcement describe the model's reasoning ability without publishing scores of this kind.
The larger sibling remains ahead: Inkling 2 Small, with 594 billion parameters, scores 72.1% on DeepSWE v1.1 against the Mini's 48.2%.
Running on home hardware
The most concrete point for local developers is speed. A Reddit user ran the model on a single RTX 4090 at roughly 190-220 tokens per second, up from about 117 in the standard setup. That required patches to vLLM, which does not yet support this card for the model, and the user reports a quality loss of about 0.4% in perplexity on some sample prompts. The setup fits in 24 GB of video memory with a 35,000-token context. The model accepts text and images as input, with a one-million-token context window.
The release follows Inkling 2 Small, which was made available on Hugging Face earlier this month. Earlier coverage is in an article on Inkling 2 Small.
The weights can be downloaded from Hugging Face. Pricing for a hosted service and the full license terms are not given in the sources consulted.

No comments yet. Be the first!