DeepSeek's V4.1 Flash update chases efficiency, not size

DeepSeek pushed out DeepSeek-V4.1-Flash-1017, a point update to its V4.1 Flash model, with weights on Hugging Face under an MIT license. The headline is not size. At roughly 763 billion parameters with a one million token context window, it sits in the same class as the version it replaces. What DeepSeek is selling is efficiency: the model card is titled "Pushing the Limits of KV Cache Compression," and the team says the update finishes agent tasks in fewer steps and with shorter outputs, which cuts the cost of each task rather than the price per token.

The benchmark jumps are large for a point release. On Terminal-Bench 4.0, a test of agent work in a real shell, the score rises from 31.2 to 42.3. ProgramBench goes from 20.3 to 29.1, and Humanity's Last Exam without tools climbs from 36.8 to 42.6. Coding and agentic scores improve, and the card reports gains on visual benchmarks too.

DeepSeek is candid about where it stands against closed models. In its own table, Anthropic's Opus 5 still leads on most tasks, scoring 51.8 on Terminal-Bench and 56.3 on Humanity's Last Exam against Flash's 42.3 and 42.6. The pitch is not that Flash wins, but that an open, MIT-licensed model now lands close enough that the gap is measured in points, not generations, and you can run it yourself. The full method is in the linked technical report.

Why it matters

If you run coding or agent workloads, the number that hits your bill is cost per finished task, not cost per token, and DeepSeek is optimizing the first one while handing you the weights. It is worth benchmarking against your current closed model before your next billing cycle.