One email a day: the top topics, each with the angle and the story, before you plan content. The daily report isn't open yet — leave your email and we'll tell you when it opens.
Image: thinkingmachines.ai
Developing
Inkling 2 Mini: Thinking Machines' 26B looped transformer for local setups
Heat 76 · #2 in AI · Rising3 reports · 3 sources · First reported Nov 21 · Updated Nov 20
The story
Thinking Machines Lab has released Inkling 2 Mini, a smaller sibling to its Inkling 2 Small model, built for running AI on local hardware. The model is a looped-transformer mixture-of-experts with about 26 billion physical parameters and 3 billion active per token: its 12 middle decoder layers run three times in a loop, so it behaves like a larger model while staying small enough to run on 16GB of RAM or a laptop. It takes image and text input, handles a 1M-token context, and ships with open weights. Hugging Face calls it state-of-the-art in its size class on the Artificial Analysis Intelligence Index v5, and says it leads in 32 of 48 benchmarks against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL. Transformers, llama.cpp, MLX and SGLang offer day-zero support, and vLLM ships day-0 support too.
day-zero support in transformers, llama.cpp, MLX and SGLang
built for local setups, runs on 16GB RAM / laptops. Hugging Face blog (second source, https://huggingface.co/blog/inkling2mini): state-of-the-art in its size class on the Artificial Analysis Intelligence Index v5, leads in 32 of 48 benchmarks against Gemma 4 31B, Muse Glimmer, Qwen3.8-27B and Nemotron 3.5 Super VL
1M-token context, image + text input, open weights
the 12 middle of 24 decoder layers run three times in a loop, so it behaves like a larger model
Nov 21
Inkling 2 Mini is a looped-transformer MoE with ~26B physical total and 3B active parameters per token
The angle
Inkling 2 Mini: the 26B model that loops its middle layers to punch above its weight - and runs on your laptopSuggested format: Short video · for Developers, AI hobbyists, local-LLM users