大模型网摘Inside V4.1-Flash-1017: How RLSF Cuts Agentic Coding Costs
DeepSeek-V4.1-Flash-1017 arrived on October 17 under the banner of "Wish-Driven Programming"—turning a one-line wish into a complete project. Based on the official release notes and API docs, this article breaks down the training method, benchmark results, and positioning.
RLSF: self-feedback as the reward signal
The core training recipe is Reinforcement Learning with Self-Feedback (RLSF): the model uses its own internal confidence scores or self-generated critiques as the reward signal for end-to-end project development, removing the need for human evaluators or external reward models. Demonstrations include 3D modeling of the Forbidden City, a web-based racing game, a Portal 2 recreation built with the Godot engine, and "Valley of Ages," a 3D scene tracing prehistoric times to the modern era. The model can also animate compact explainer videos.
Reading the DeepSWE numbers
| Model / setting | DeepSWE v1.1 Pass@1 | Avg cost per task |
|---|---|---|
| V4.1-Flash-1017 | 77.0% | ~$0.20 |
| V4.1-Flash-1017 (low) | 72%+ | ~$0.10 |
| V4.1-Flash (predecessor) | 74.2% | ~$0.50 |
The cost drop comes largely from shorter outputs: agentic coding costs are dominated by output tokens, and the official chart shows several larger models spending $2–$10+ per task. Note these are vendor benchmarks; see the comparison article for context.
Specs and family positioning
On the API, deepseek-flash now points to V4.1-Flash-1017: 1M context, 384K max output, vision input, JSON output and tool calls, thinking mode on by default, and a 2,500 concurrency limit. The parallel deepseek-v4-pro (V4-Pro-0813) shares the context window but has no vision and a 500 concurrency cap. For the September-era positioning, see the V4.1-Flash profile. The release cadence—V4 Preview (April), V4-Pro GA (August), V4.1-Flash (September), and this 1017 update—keeps accelerating.
How to get it
Weights are open on Hugging Face (deepseek-ai/DeepSeek-V4.1-Flash-1017) and ModelScope; the API serves it under the deepseek-flash name with legacy names auto-mapped. DeepSeek recommends Harness v0.2.1-rc.1 or later—see the tools roundup—and the freshly opened Search & Fetch APIs are covered here.