AMD's Instella-MoE: an open model that ships with its full training recipe
AMD released Instella-MoE-16B-A3B, a mixture-of-experts model with 16 billion total parameters that activates only 2.8 billion per token, trained end to end on its own Instinct MI300X and MI325X GPUs. Unusually, the release includes checkpoints from every training stage plus the data mixtures, training configs and code — the code under MIT, the weights under a research-only license. AMD says the base model posts the strongest average among fully open models, ahead of Moonlight-16B and OLMo-3.
Why it mattersMost 'open' models ship weights and nothing else. A release that includes the whole training recipe lets researchers actually study and reproduce how a model is built — and it is a working proof that competitive training runs can happen entirely off Nvidia hardware.
▶ Related video: AMD Releases First Ever AI model: Instella-MoE-16B-A3B-Think
Read in the app — free, in 9 languages
Related stories
DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best2026-08-21Show the robot once — three to twelve seconds — and it gets the job right 59 times out of 100 with no training at all
2026-08-21The model invents its own tasks, builds the rig to test them, then trains on the results — and DeepReinforce gave the weights away
2026-08-20A free model that fits on one graphics card scores 52 — then spends 22,276 thinking tokens drawing a picture
2026-08-18Zhipu did not build a new model. It kept training the old one — and says coding got 50 percent better.
2026-08-17