Thinking Machines releases Inkling 2 Mini: 26B looped MoE with 3B active params, 1M context
mervenoyann · x · 2026-11-21
Thinking Machines Lab has released Inkling 2 Mini, a looped transformer built for local setups: a multimodal Mixture-of-Experts model with 26B total parameters and only 3B active per token. Its capacity trick is visiting some layers in a loop that repeats three times, so it behaves like a much larger model. It supports 1M-token context and accepts image and text inputs.
- Ships in mixed precision with NVFP4 and MXFP8 expert weights
- Includes a DSpark drafter for speculative decoding speed-ups at small memory cost
- The Hugging Face team vibe-tested it as a coding companion and shipped support in transformers, llama.cpp, MLX and SGLang
Official benchmark charts show it leading same-class rivals like Gemma 4 31B and Qwen3.8-27B on AA Intelligence Index v5.0, Terminal-Bench 4.0, FrontierCode 2.0 and DeepSWE v1.1.
More from Infra
- Ollama CEO in Fortune: enterprise demand for open models keeps growing — ycombinator · 2026-11-21
- Nvidia plunges 13.17% in a single day to $126.75 — jfischoff · 2026-11-21
- vLLM ships Day-0 support for TML model, cuts GB200 decode latency to 3.84ms — vllm_project · 2026-11-21
- vLLM explains looped Inkling internals: shared weights with separate per-pass KV state — vllm_project · 2026-11-21
- vLLM ships day-0 support for Thinking Machines' new looped Inkling model — vllm_project · 2026-11-21
- Starcloud built a $2.5M space data center with the first H100 in orbit — chetanp · 2026-11-21