I m not sure I would describe this as blanket win over GPT-5.6 Sol. In DeepSeek’s own table, V4.1 Flash is ahead on Terminal-Bench 2.1, DeepSWE, NL2Repo, and AutomationBench, but it is behind on GPQA Diamond, Terminal-Bench 3.0 and 4.0, and SEC-Bench Pro.
The architecture is probably part of the explanation for the lower cost and faster inference. DeepSeek says V4.1 Flash uses a new Causal Encoder–Decoder design, with 8B active parameters for input processing and 16B for decoding, along with much smaller KV caches.
But I hope it is just not benchmaxxed and genuinely good model
It is pretty funny that they recently announced about mythos which possess cybersecurity threat and then after some days, the claude code leaked. I think we know the culprit
The architecture is probably part of the explanation for the lower cost and faster inference. DeepSeek says V4.1 Flash uses a new Causal Encoder–Decoder design, with 8B active parameters for input processing and 16B for decoding, along with much smaller KV caches.
But I hope it is just not benchmaxxed and genuinely good model
benchmarks: https://media2url.com/m/52a77a33347c48