Synth/agent pipelines aside this is the main catch of DeepSeek 3.2: "This isn’t a “new model card, new Elo” update. It’s an architectural update". At mid-training scale, everything is mutable and we are likely to see more and more sub-model speciation.
Loading thread detail
Fetching the original tweets from X for a clean reading view.
Hang tight—this usually only takes a few seconds.