Today we’re releasing the weights for Laguna M.1,
our most capable model to date, with a 256K context length.
Both base and post-trained checkpoints are now available on Hugging Face under Apache 2.0.
@alexocheema Ohhhh sorry, to clarify, it was in response to you quoting the GLM-5.2 release! GLM-5.2 kinda looking like we might finally have an Opus 4.5 level open weights model
In case you even had a single thought about trying the "Qwable" local AI model - the Mythos / Fable "Distillation" / Finetune of @Alibaba_Qwen Qwen3.6-35b MoE... https://t.co/7VAwCm502N
@Zai_org Dropped the weights for GLM-5.2, probably the best open model in existence right now. Not that most of us have the hardware to run it - https://t.co/kIS1ONA6ka
Thank you @RyanLeeMiniMax@MiniMax_AI for the API credits to test MiniMax M3! A fantastic open weights model that gives me a new /goal: accumulate more hardware to run it locally!! https://t.co/kPkiILVXgc
TLDR: AI will grow the economy. The new Intellectual Property (IP) each company protects will be their use of AI; how it's implemented, maintained, and improved
@chongdashu@DevOuterReaches recommended Qwen3.6-27B and while it's legit an awesome model, I promise you'll want to choose a MoE instead of a dense model, it will be much faster. Compare it to Qwen3.6-35B-A3B to see the speed difference. With 128gb you could also run larger MoE models
Well, with Fable pulled, figured I should dust off the old @Zai_org coding plan and test out GLM-5.2. Oh, and also compare it to @MiniMax_AI M3 - https://t.co/2PQ69mJY2M
@notallthere_net@0xajka@theo I use Codex every day. The point is that it has been made clear the government can shut down any US lab, at any time. Local models are not as good, but they also aren't years behind, and are VERY capable at both coding and agentic orchestration and automation