
Tech • AI • Robotics • Game
⏱ ≈ 3 min read · 21-min video — 18 min saved
Thinking Machines has released Inkling 2 Mini, a 26 billion-parameter open-weight multimodal model that emphasizes local efficiency, posting strong coding results while running on consumer hardware with only 5.44 billion active parameters per token.
Inkling 2 Mini uses a looped transformer design that repeatedly routes information through middle layers, increasing reasoning depth without increasing parameter count. Although the model has 26 billion total parameters, only 5.44 billion active parameters are used per token across its reasoning loops. On Artificial Analysis Intelligence Index, it scored 20.1, placing it on the Pareto frontier for intelligence relative to active parameters.
The model is released under the Apache 2.0 license and supports both text and image inputs. It also offers a 1 million-token context window, a rare specification for a model positioned for local use. Thinking Machines is also releasing a variable loop research model to study how additional reasoning passes affect performance and transparency.
A central selling point is that the model can run locally on systems such as a 32 GB RAM Mac, including via Ollama, and on devices like Nvidia DGX Spark. Reports also indicate it has been run on laptops, and an iPhone version is said to be in development. Quantized versions as small as 5.4 GB and 8.9 GB reportedly retain 91.2% and 96.8% of original performance.
Local speed figures are unusually high for this class. The model reportedly reaches 210 tokens per second on an M5 Ultra Mac Studio and 125 tokens per second on an M5 Max MacBook Pro, with prefill speeds of 3,620 and 1,750 tokens per second respectively. Performance of around 200 tokens per second has also been reported on an RTX 5090.
Benchmark and practical testing place Inkling 2 Mini ahead of Qwen 3 27B on several coding-heavy tasks, especially agentic coding, tool use, and local software generation. Reported scores include 48.2% on DeepSWE, 77.9% on SWE Bench Pro, and 79% on BrowserComp. In one coding test, it built a native weather app 1.8 times faster than Claude Sonnet 5.5 while being 11 times cheaper on API pricing.
Independent evaluation cited in the release ranked the model number one in efficiency and intelligence among open-weight models in that test set, and ninth overall in intelligence with a score of 78.4. It also posted a lower cost per correct answer of $25.80. Researchers at Stanford also found it significantly more energy efficient than comparable local models.
The model is not presented as universally superior. It reportedly trails Qwen 3 27B in multi-turn conversation and complex instruction following, while larger frontier models still outperform it on harder reasoning tasks. Knowledge-heavy tasks also remain an area where competing systems keep an advantage.
Testing highlighted strong results in web development and interactive front-end generation, including 3D scenes, SVG graphics, dashboards, small games, and product viewers. In multiple side-by-side tasks, the model often completed jobs faster and with fewer output tokens than Qwen 3 27B, while producing more polished or functional code. Not every comparison favored Inkling, but it was consistently competitive and often more concise.
Inkling 2 Mini appears to mark a notable step forward for local AI, combining open weights, long context, multimodal input, and unusually high efficiency in a package that can run on mainstream hardware. Its main significance lies not in topping every benchmark, but in bringing strong coding performance and practical usability to devices far below data-center scale.
Ask a question