DeepSeek V4.1 Flash hands-on tests and comparisons

The post compares DeepSWE 77.0, Terminal-Bench 42.3 and Automation-Bench 60.5 against Sol and Opus 5, alongside per-million-token pricing for Flash, Luna, Sol and Opus 5.

A daily AI roundup notes V4.1-Flash-1017 is live with 763.2B parameters tuned for coding harnesses, alongside items on Reflection AI's Beam, a China draft on open-weight export controls, and MAI-Base-2.

A single prompt produced a Call of Duty-style Three.js FPS with 5 survival waves, 3 operators and 3 difficulties in 37 minutes for about $0.10; the author rates it on par with Gemini 4.

Long post breaks down the launch: DeepSWE 77.0 over Opus 5, Agents' Last Exam 35.4 vs 28.6 but 14 points behind on HLE, $0.003 per million cached input tokens, plus open Search & Fetch APIs.

Comparison take: the author finds V4.1 Flash-1017 better than GPT-5.6-SOL and questions the value of a $200 Codex subscription.

The new Harness builds add V4.1-Flash-1017 support, DeepSeek's own Search and Fetch APIs, custom plugin indexes, hot reload and experimental plugins, with breaking changes for plugin authors.

The poster says V4.1 Flash 1017 runs on one RTX 5090 plus CPU offload as the fastest setup for that hardware, without giving tok/s figures.

A roundup of the launch: web, app and API availability, MIT weights on Hugging Face, 2.0–11.1 point gains across 11 benchmarks, 77.0 on DeepSWE ahead of GLM-5.3 and Opus 5, plus tuning for Harness and Tencent CodeBuddy.

The post says V4.1 Flash moved to the 1017 build across web, app and API, opening self-built Search and Fetch APIs at ¥5 per 1,000 calls, callable server-side from Claude Code and Codex, free before November 1.

Deployment breakdown: offloading the ~203GB Engram table to SSD and quantizing the rest makes 256GB setups viable, following prior 2-bit Mac Studio and dual DGX Spark EXL3 recipes.

A user posts video of DeepSeek-V4.1-Flash-1017 running in the official DeepSeek Harness.

The Oct 17 brief covers Gemini 4 Argon in AI Studio Build for AI Pro users and in Perplexity for Pro/Max, plus GPT-6 Astra Pro and DeepSeek V4.1 Flash news.