Thinking Machines Lab has released Inkling 2 Mini, a multimodal mixture-of-experts model that it says delivers strong performance on hardware small enough for local setups. The model has about 26 billion total parameters, of which roughly 3 billion are active for each token. It accepts text and image inputs and supports a context window of up to 1 million tokens.
The model's main design feature is a looped architecture. It has 24 physical decoder layers, but the 12 middle layers run three times, so each token passes through 48 logical layer visits. The passes reuse the same weights, and a gating mechanism lets each loop pass refine the output of the previous one. The company says this gives the model more effective capacity than its stored size suggests. It ships in a mixed-precision format with NVFP4 and MXFP8 expert weights, and includes a DSpark drafter for faster generation at a small memory cost. A research variant that can run with a variable number of loop passes is also available. Using one or two passes instead of three speeds up inference at some cost in output quality.
The company reports that Inkling 2 Mini leads 32 of 48 benchmarks against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL, and is state-of-the-art in its class on the Artificial Analysis Intelligence Index v5. These results are self-reported, and the company says more independent validation is needed. Its informal tests placed it at least on par with Qwen 3.8 27B. The model is supported in transformers, llama.cpp and MLX for local use, and in SGLang for datacenter deployment. For Australian businesses and professionals, the practical impact is limited for now. The weights can be downloaded without per-token API fees, but running a model of this size needs a capable GPU or high-memory Apple silicon machine. Keeping prompts on local hardware may simplify privacy compliance under the Privacy Act, though it does not remove obligations. Teams should test it on their own tasks before relying on the published benchmarks.
Key Points
Inkling 2 Mini is a multimodal mixture-of-experts model with about 26B total and 3B active parameters per token.
A looped architecture runs the middle 12 of 24 physical layers three times, giving 48 logical layer visits per token.
The model supports a 1M-token context window and accepts text and image inputs.
Thinking Machines Lab reports it leads 32 of 48 benchmarks against several rival open models, though results are self-reported.
Transformers, llama.cpp and MLX support local use, while SGLang targets datacenter deployment.