AI Agent News/stories/thinking-machines-lab-releases-inkling-2-mini-for-local-setups.md

Infrastructure

Thinking Machines Lab releases Inkling 2 Mini for local setups

Thinking Machines Lab has released Inkling 2 Mini, a multimodal mixture-of-experts model with about 26B total and 3B active parameters per token, aimed at local deployments. Hugging Face says it is supported in transformers, llama.cpp, MLX and SGLang from day zero.

Source: huggingface.co/blog/inkling2mini

Inkling 2 Mini is a multimodal mixture-of-experts model with about 26B total and 3B active parameters per token. It accepts image and text input and supports a 1M-token context. Its middle layers run in a loop three times, so 24 physical layers act as 48 logical ones without storing that many distinct layers.

Hugging Face 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. The model ships with NVFP4 and MXFP8 expert weights and a DSpark drafter for speculative decoding. The authors say more validation is needed after their informal vibe tests, which found it at least comparable with Qwen 3.8 27B.

Inference support covers transformers, Llama App and llama.cpp with GGUF 4-bit and 8-bit quants, MLX, and SGLang. A research variant can run one or two loop passes instead of three for faster generation, at a possible cost in quality.

Why it matters if you build agents

Developers running agents on local hardware now have a open-weight model that the publishers position as a strong coding and knowledge-work option, with support across common local stacks. Its reasoning effort can be set per request, which lets builders trade latency against quality. Benchmark claims come from the publishers and their own tests, so teams should evaluate it on their own agent tasks before relying on it.

Get stories like this every weekday

One short email with the day's agent tooling news.

Subscribe