DeepSeek Releases V4.1-Flash-1017 with Enhanced Performance and Lower Costs
nashnova research
DeepSeek launched its V4.1-Flash-1017 model on October 17, boosting code and agent-task benchmarks while cutting run time and output length — what the company calls a Pareto improvement, better and cheaper at the same time.
What exactly got upgraded?
V4.1-Flash-1017 scores higher across code generation and agent-task benchmarks, while shortening both run time and output length.
This means → the model is not just more accurate — it is faster and leaner per call, so each API request costs less.
DeepSeek labels this a "Pareto improvement" — a gain where no metric gets worse. In plain terms = they didn't trade performance for price; both improved.
What is "wish-based programming"?
DeepSeek introduced wish-based programming: a user describes a goal in one sentence, and the model handles planning, coding, running, checking, and fixing on its own.
The feature runs on self-feedback reinforcement learning — the model generates a first draft, then keeps self-reviewing and iterating without step-by-step human guidance.
Official demos include a browsable 3D Forbidden City scene, a browser racing game, a Portal 2 remake, and a dynamic 3D world spanning prehistory to the modern era.
How do agent frameworks perform?
On the DeepSWE benchmark, the new model pushed both the performance and cost frontiers across multiple agent frameworks.
DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy all received targeted optimizations; using the new model requires upgrading DeepSeek Harness to v0.2.1-rc.1 or later.
This reflects a shift from "standalone model power" to framework-plus-model co-tuning — tighter toolchain integration lowers the barrier for developers.
How cheap is the new search API?
DeepSeek simultaneously launched two self-built search-engine APIs: Search (web search) and Fetch (extracts page content into Markdown).
Pricing sits at ¥5 per 1,000 successful requests — roughly half a cent per call; failed requests, empty results, and timeouts are free.
The APIs work inside DeepSeek Harness, Claude Code, or Codex, and are free until November 1. In plain terms = free now, near-free after.
What comes next?
Hours before the official launch, users spotted the new model briefly appearing in the model list, and the system fingerprint had already changed — the backend switch preceded the announcement.
Early user feedback converges on three points: smarter, faster, more concise output, with improved 3D generation.
Whether the model can keep widening its cost-performance lead in the agent-coding race depends on the pace of future updates and how deeply the developer ecosystem actually adopts it.
市场有风险，内容仅供研究参考，不构成投资建议。