DeepSeek released DeepSeek-V4.1-Flash-1017 — a pinpoint update to its fast open model. The company promises not only higher scores in reasoning, code, agents and images, but shorter answers: the job finishes faster and cheaper.
What changed in the numbers
On Terminal-Bench 4.0 the model scores 42.3 against 31.2 for the previous Flash. DeepSWE v1.1 — 77.0 after 74.2. ProgramBench jumps from 20.3 to 29.1, Apex Bench from 64.6 to 72.1. On Humanity’s Last Exam without tools — 42.6 against 36.8. Images with tools rose too: Chartography 83.0, BabyVision 92.5. This is the vendor’s table, not an independent run, but the direction is one: agent code and vision, not chatter.
The repo lists about 763 billion parameters and an MIT licence. Context is a million tokens, default sampling temperature 1.0. There is no separate Jinja template: prompt encoding moved into a Python reference and the Rust deepseek-recipe kit.
Why cut answer length
Most model “improvements” inflate thinking tokens. DeepSeek is clearly moving the Pareto line: less text for the same or better accuracy. For an agent that drives a terminal, a shorter answer means a smaller bill and less chance of drowning in its own draft.
How to look at it
This is not a new architecture, it is a patch to Flash. It matters if you already run V4.1 locally or through OpenRouter and hit the cost of long chains. Comparing it to closed Opus on one table row is pointless: the harnesses differ. Watch whether the gain repeats on your harness, not on DeepSeek’s slide.
