DeepSeek V4.1 Flash hands-on tests and comparisons

Identical config.json and 16B active params to the September build; DeepSWE hits 77% at ~$0.20 per task versus 74% at $0.50 before, and 72.5% at $0.11 in low effort.

A rundown of vendor-chart figures: MIT-licensed open weights, 77% on their own DeepSWE, about $0.20 per task, with a caveat to await independent benchmarks.

Niels Rogge shares the Hugging Face release: DeepSeek-V4.1-Flash-1017 is a minor update to V4.1-Flash with substantial gains in reasoning, coding, agentic tasks and visual benchmarks.

A blogger condenses DeepSeek's official update into one image: no price increase and across-the-board capability gains, linking the official announcement.

A Harness test generating a DeepSeek whale promo video with SVG, Canvas and Three.js: 247 tok/s, 49.4M tokens, 99% cache hit, total cost ¥2.17.

Using DSH plus the API, DeepSeek-V4.1-Flash-1017 reviewed recent Opus 5.5 commits, found one P0 and several P1 issues that Opus 5.5 confirmed, all for about ¥0.3.

Per its model card, V4.1-Flash-1017 improves reasoning, coding, agentic and visual benchmarks while cutting execution time and output length; Search and Fetch APIs open at $0.75 per 1,000 calls each, with free weights on Hugging Face and ModelScope.

The model card reports Terminal-Bench 4.0 at 42.3, ProgramBench at 29.1 and DeepSWE at 77.0, beating GPT-5.6 Sol and Opus 5, though HLE and Terminal-Bench still trail Opus 5; web, app and API are live, weights are on Hugging Face, and it is tuned for DeepSeek Harness and CodeBuddy.

Tested with the prompt "create svg of red lamborghini", the 16B-active Flash model produced full SVG geometry on the first try, which the author found surprising.

Official update: Search and Fetch APIs open to developers, V4.1-Flash moves to 1017 at unchanged pricing, with faster tasks, shorter outputs and DeepSeek Harness tuning.

vLLM reports DeepSeek-V4.1-Flash serving 830K tok/s per GPU and 1.3B tokens per dollar TCO on SemiAnalysis AgentX, with 5x agentic throughput since day one.

DeepSWE 77.0 vs 74.0, Automation-Bench 60.5 vs 50.3, Agents' Last Exam 35.4 vs 28.6 and Apex 72.1 vs 65.8, though Terminal-Bench and HLE still trail; the post stresses this is the smallest 552B MoE model with open weights.