The development
What happened
Hugging Face’s 20 November release account introduces Thinking Machines Lab’s Inkling 2 Mini, a multimodal mixture-of-experts model aimed at local AI setups. It accepts images and text and advertises a one-million-token context window. About 26 billion physical parameters and three billion active parameters per token are different from a three-billion-parameter checkpoint: the design reuses a core of layers, while still performing repeated computation and retaining per-visit state. Local inference integrations accompany the release, but hardware fit, performance and the separate variable-loop research checkpoint need their own testing.
Key facts
The model has 24 physical decoder layers. It visits layers 0–5 once, layers 6–17 three times and layers 18–23 once: 48 layer visits per token, with the core weights reused.
Cross-pass messages and learned gates let later passes refine earlier computation. Each layer visit still keeps its own key/value and short-convolution state; fewer distinct weights do not eliminate that work.
The shipped expert weights mix NVFP4 and MXFP8. Hugging Face also describes an optional DSpark drafter for speculative decoding, adding a memory-versus-speed tradeoff rather than a free acceleration guarantee.
The release account announces Transformers, Llama App/llama.cpp, MLX and SGLang support. GGUF options include 4-bit IQ4_NL and 8-bit Q8_0; the documented SGLang example requires SM103 hardware.
A separate TML-d24-varloop-research checkpoint allows fewer loop passes. Hugging Face says it tested that variant but has not evaluated it in depth; faster inference can come with lower output quality.
Hugging Face reports Inkling 2 Mini ahead of Qwen3.8-27B on FrontierCode 2.0, tied on Terminal-Bench 4.0 and behind on AutomationBench-AA 1.0.6. These are reported comparisons, not TideReach’s independent benchmarks.
The context
Why it matters
TideReach assessment: Inkling 2 Mini makes weight reuse a concrete local-model design choice, rather than simply shrinking every layer. Teams should distinguish checkpoint size, active parameters, repeated computation and context-state memory when deciding what fits a machine. The inference integrations lower the setup barrier, but neither a one-million-token context claim nor a successful demo establishes affordable full-context operation, general agent reliability or a security guarantee. TideReach has not installed or benchmarked the model.