Mistral AI explains how it built Mistral Large 4: on long tasks it squeezes the model's memory less and reads only the most relevant parts
Models often shrink their memory of past text to save cost. Mistral AI found an early Mistral Large 4 with memory cut to 1/16 held up on knowledge tests but got worse at agent tasks. The final model gives more layers full memory and reads only the most relevant parts. Of its 1T parameters, 52B handle each word, and a 606 GB lower-precision copy runs on 4 GB300 GPUs. Some layers still share memory.
Why it matters
A warning for teams building long-running agents: a memory-saving change can look fine on standard tests yet hurt the real task, so test model designs on the work they are meant to do.
Sources1
Written by dalao.ai: the title, summary and why it matters are our own words. The original reports are the record; their publishers keep authorship.