DeepSeek Just Made Reasoning WAY More Efficient!

DeepSeek calls V4.1 Flash 1017 a “minor update”, but it went up on every benchmark in their table, and it’s now the #1 open model on Arena’s WebDev leaderboard. The bigger story is tokens. DeepSeek has always been cheap per token, but its Flash models were token-hungry, so cheap per token never meant cheap per task. In my reruns, 1017 used about a third fewer thinking tokens at high effort and 45% fewer at max, with the same prompts and settings. 🔗 Links DeepSeek announcement: https://x.com/deepseek_ai/status/2111375640717033777 Model card (Hugging Face): https://huggingface.co/deepseek-ai/