DeepSeek Just Released V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API

MarkTechPost reports DeepSeek released DeepSeek-V4.1-Flash-1017, a new post-training checkpoint of V4.1-Flash live on DeepSeek’s website, app, and API as deepseek-flash. DeepSeek calls it a Pareto improvement over the September release — stronger scores, faster task completion, shorter outputs — and opened Search and Fetch APIs to every developer.

Per DeepSeek’s figures cited by MarkTechPost, the model beats prior V4.1-Flash on all 11 reported benchmarks by 2.0 to 11.1 points, with a best of 77.0 on DeepSWE v1.1 (above Claude Opus 5’s 74.0 in DeepSeek’s table) and a weaker 42.6 on HLE without tools versus Opus 5’s 56.3. Architecture notes from the V4.1-Flash report: 552B backbone plus 196B Engram parameters, activating 8B per token in prefill and 16B in decode, 1M context. September V4.1-Flash weights remain MIT-licensed on Hugging Face; 1017 weights and hardware requirements were not disclosed.

DeepSeek is selling the agent loop as a bundle: a fast coding model plus first-party search behind one API key. That’s practical for builders who’ve been stitching OpenAI-style chat to a separate serper or browser tool.

We’ll treat DeepSeek’s leaderboard as company numbers until independent harnesses reproduce DeepSWE 77.0. The pattern still matters — open Chinese labs keep shipping sparse, long-context coding models that pressure Western pricing. Opening Search/Fetch to every developer is the quieter product move.

Useful for agent teams on a budget. The downside is dependence on DeepSeek’s cloud if 1017 weights stay closed. I’m watching whether the MIT September weights get a matching open 1017 drop, and how Terminal-Bench gaps versus Opus 5 play out in real harnesses.

DeepSeek says 1017 uses reinforcement learning with self-feedback for end-to-end project work and is tuned for DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy.