DeepSeek-V4-Flash edges Qwen3.6 on benchmarks

The chart gives DeepSeek the headline win, but the real story is the same one the comments point to: benchmarks flatten fast, while latency, quantization behavior, and context scaling decide what people actually keep using.

• –DeepSeek-V4-Flash is a preview model with 284B total parameters, 13B active, and a 1M-token context, so it is competing on capability density plus long-context headroom
• –Qwen3.6 is the more practical local-first family in this debate, especially the 27B dense and 35B-A3B open models that are easier to serve and tune
• –Reddit commenters are skeptical of the graph itself, arguing that small benchmark deltas can hide much larger real-world differences in knowledge and tool use
• –The discussion also highlights a key MoE tradeoff: sparse models can look cheap on active params, but real speed still depends on architecture, cache behavior, and context length
• –For builders, this is less a “winner” announcement than a reminder to test coding, retrieval, and agent loops on your own workloads before switching models