DeepSeek Releases V4.1-Flash-1017, Updating Open Weights and Coding Agent Benchmarks

DeepSeek released DeepSeek-V4.1-Flash-1017 on October 17, updating the existing V4.1 Flash to a new snapshot. According to the official changelog, API calls and pricing remain unchanged, and using deepseek-flash will return the new version. Applications that use this name should therefore recheck their outputs and tool-calling behavior. Official changelog

The update focuses on the ability of coding agents to complete tasks. According to the official model card, Terminal-Bench 4.0 rose from 31.2 to 42.3, an increase of 11.1 percentage points; DeepSWE v1.1 rose from 74.2 to 77.0, while ProgramBench increased from 20.3 to 29.1. The model card notes that the first two benchmarks use the Claude Code and Mini-SWE-V2 execution frameworks, respectively. Their results should be understood in the context of the tool environment and execution settings, and cannot be taken as gains that every agent product will achieve. Official model card

DeepSeek says the new version uses reinforcement learning with self-feedback to improve end-to-end project development, while reducing task time and output length. The company recommends using DeepSeek Harness v0.2.1-rc.1 or later. For engineering teams, the key metric to validate is the total cost per successful task: shorter outputs may not reduce overall costs if they come with more retries. For now, these efficiency improvements remain results reported by the provider. Release announcement

Self-hosting also involves specific integration requirements. The weights are licensed under MIT, but the release package does not include a Jinja-format chat template. Instead, it provides a Python reference implementation for prompt encoding, as well as deepseek-recipe, a Rust utility library with Python bindings. The library can convert between multiple API request formats and handle reasoning content, tool calls, and images. Inference, tool execution, and HTTP transport are still the responsibility of the caller, so deployers should check that their existing services preserve these semantics correctly. Encoding and deployment guide

The next thing to watch is whether independent testing can reproduce the gains in success rate and task time, and whether the model performs similarly on Chinese-language tasks, multi-turn tool use, and modifications to existing codebases. In practice, teams can compare completion rates, total tokens, and end-to-end time using a fixed task set, execution framework, and retry budget, rather than deciding whether to migrate based on a single leaderboard result.