Why DeepSeek V4.1 Flash feels different on OpenCode: serving, not weights
The author argues quantization isn't the main factor; KV cache policy, routing and serving optimizations matter more, with speeds ranging down to 3 tok/s and aggressive context compaction on free tiers.
i’ve been looking more closely at why DeepSeek V4.1 Flash feels different on OpenCode. 5 things stood out: 1. i don’t think quantization explains everything. V4.1 Flash already ships in MXFP4/FP8 mixed precision, so there isn’t that much room left to keep squeezing it. 2. cache behavior looks like a much bigger deal. DeepSeek’s own infra reportedly keeps KV cache around much longer and shares it across more of the serving stack. 3. speed is all over the place. i’ve seen reports ranging from normal performance down to 3 tok/s, which makes provider quality matter a lot more than the model name. 4. the free version also seems to compact much more aggressively for some people, which can make the model feel worse even if the underlying weights are identical. 5. i care more about transparency now. if OpenCode is changing kernels, cache policy, routing or serving optimizations, i’d rather know that than assume “DeepSeek is just worse here.” my takeaway: the same model can feel completely different depending on how it’s served.