Thinking Machines Lab releases Inkling 2 Mini, a ~26B-parameter (3B active) looped mixture-of-experts multimodal model for local use, with Hugging Face transformers, llama.cpp, MLX and SGLang support
Edited by Elliot Marsh, Editor · written with AI from the linked sources (how)
Thinking Machines Lab published open weights for Inkling 2 Mini (thinkingmachines/Inkling-2-Mini), a multimodal (text+image) MoE with ~26B total and 3B active parameters, 24 physical layers (middle 12 looped 3x for 48 logical layer visits), 128 experts with 8 routed per token, 1M-token context, mixed NVFP4/MXFP8 expert weights, and a DSpark drafter for speculative decoding. A research variant (TML-d24-varloop-research) accepts a variable number of loop passes. Day-zero support is available in transformers, llama.cpp, MLX and SGLang.
Entities: Thinking Machines Lab, Hugging Face, Inkling 2 Mini, Inkling 2 Small, Qwen3.8-27B, Gemma 4 31B
1 primary
What happened
Thinking Machines Lab has published open weights for Inkling 2 Mini, a text-and-image model with about 26B total parameters, of which 3B are active for each token. It uses 24 physical layers, with the middle 12 reused three times (48 layer passes), 128 experts with 8 chosen per token, a 1M-token context window and mixed NVFP4/MXFP8 expert weights. A speculative-decoding drafter is included, and a research variant allows a variable number of loop passes. Support for transformers, llama.cpp, MLX and SGLang was available on day one, so people can download and run it now.
Why it matters
Developers and researchers who run models locally get another credible option in the roughly 27B class. Only 3B parameters are active per token, so it should be cheaper to run than a dense model of similar size, though all 26B of weights still have to fit in memory. Broad runtime support lowers the cost of trying it, so teams can test it against Qwen or Gemma on their own tasks this week. The looped-layer design is also an architecture experiment worth watching, but its real-world value is unproven.
What is noise
The headline claims (leading 32 of 48 benchmarks, topping the Artificial Analysis index for its class, "state-of-the-art") come from the lab's own launch material, and the comparison set was chosen by the vendor. Hugging Face's own informal tests only say it is "at least comparable" to Qwen 3.8 27B and that more validation is needed, which is much weaker than a clear win. The 1M-token context is a stated capability, not evidence that quality holds up at that length.
Watch next
- 01Independent evaluations (Artificial Analysis re-runs, community leaderboards, Hugging Face's follow-up validation) showing whether the lead over Qwen 3.8 27B and Gemma 4 31B holds on tasks outside the vendor's chosen 48 benchmarks.
- 02Real-world tokens-per-second and memory figures on consumer hardware (llama.cpp, MLX, SGLang), including whether the looped layers and drafter deliver the promised speed, and how quality behaves at long context.
- 03Downloads, fine-tunes and quantisations on Hugging Face over the next few weeks, plus any release of Inkling 2 Small or published results for the variable-loop research variant.
Coverage
1 storyNewsletter
Get the no-hype AI news straight to your inbox
Weekly. No spam, unsubscribe any time. Privacy
More capability signals
Full feed →- Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents10 Sept 202682
- Mistral releases Mistral Large 4, a ~1T-parameter open-weight multimodal MoE model, with day-0 Hugging Face support29 Oct 202680
- Meta releases Muse Spark Open, an open-weight ~816B-parameter multimodal mixture-of-experts model, with Hugging Face support15 Oct 202680
- OpenAI discloses sandbox-escape and credential-leak incidents, confirms pause on tool-use for its most capable models26 Sept 202680