DeepSeek updates V4.1 Flash with gains aimed at coding agents

DeepSeek has released DeepSeek-V4.1-Flash-1017, a minor update to its V4.1 Flash model that the company says performs better on reasoning, coding, agent and visual tasks. DeepSeek says the update also completes tasks faster and produces shorter outputs. The model is available through DeepSeek’s website, app and API, and its weights can be downloaded from Hugging Face and ModelScope.

The model card’s results show gains over the previous V4.1 Flash in all 11 listed benchmark rows. The largest jumps include Terminal-Bench 4.0, which rises from 31.2 to 42.3, and ProgramBench, from 20.3 to 29.1. DeepSeek reports a smaller increase on DeepSWE v1.1, from 74.2 to 77.0. These are company-published comparisons, not independent evaluations; the card specifies Claude Code for Terminal-Bench and Mini-SWE-V2 for DeepSWE.

Against the other models in DeepSeek’s scorecard, the results are mixed. V4.1-Flash-1017 scores higher than the listed Opus 5 reference on DeepSWE, Automation-Bench, Agents’ Last Exam and Apex Bench, but lower on the other seven benchmarks. DeepSeek’s visual benchmark rows use tools, so those scores reflect a model-and-tools setup rather than an unassisted model.

DeepSeek is also emphasizing end-to-end project work. It says reinforcement learning with self-feedback helped the model build projects from scratch, pointing to examples including a web racing game, a Godot recreation of Portal 2 and 3D scenes. These are demonstrations of selected outputs, not evidence of a general success rate for building projects.

The release is aimed in part at developers using agent harnesses: DeepSeek says the model is optimized for DeepSeek Harness, Tencent CodeBuddy, WorkBuddy and other environments, and recommends DeepSeek Harness v0.2.1-rc.1 or later. The model card lists a 1-million-token context window and provides local inference instructions, along with MIT-licensed weights. DeepSeek also released prompt-encoding code and a Rust toolkit with Python bindings; users must supply their own inference, tool execution and HTTP transport.

Alongside the model, DeepSeek opened its Search and Fetch APIs to all users, pricing each at $0.75 per 1,000 calls. Server-side calls through DeepSeek Harness, Claude Code and Codex are free for DeepSeek API users for now; the company says billing for those calls begins October 31, 2026, at 16:00 UTC. DeepSeek’s launch materials do not establish a universal cost saving from the model’s shorter outputs or faster completion, so the practical economics will depend on the workload and tool use.