DeepSeek V4.1 Flash hands-on tests and comparisons

A Japanese post summarizes the release: 77.0 on the DeepSWE v1.1 code-repair benchmark versus Opus 5's 74.0, Search/Fetch APIs at $0.75 per 1,000 calls, MIT-licensed weights, 1M context and image understanding.

A Japanese quick report: released October 17, available on web, app and API under the name deepseek-flash, with better performance and shorter output suited to iterative game prototyping.

Argues that with a 91.5% cache hit rate the effective input cost is about $0.017 per million tokens, alongside the $0.75 per 1,000 calls Search/Fetch API and 552B MIT-licensed weights.

On Noahbench Homework v1, V4.1 Flash 1017 scores 55% for 5th place, tying Muse Wonder while costing $0.81 versus $6.33, at the price of 1h18m for 20 tasks.

Summarizes the release: MIT-licensed weights on Hugging Face and ModelScope at unchanged pricing, a self-reported 77.0% on DeepSWE v1.1, and one-prompt demos like a 3D Forbidden City and a browser racing game.

A user calls it the best value among domestic tools; DeepSeek's Search and Fetch APIs cost 5 yuan per 1,000 successful calls, and in-tool search in Claude Code, Codex and DeepSeek Harness starts billing on Nov 1.

On the same parser coding task both scored 30/30; Flash took 88 seconds and $0.0125 versus Pro's 20.7 minutes and $0.14, with $0.055 vs $0.32 across the first four tests.

Running deepseek-v4.1-flash:1017 on Ollama Cloud gave about 220 output tok/s sustained and a 0.25–0.6 second time to first token.

A user gave a reference video and asked the model to recreate multi-scene 3D traversal with Sun Wukong in Blender; the result fell short, a concrete failure case.

The author says V4.1 Flash is roughly on par with Kimi K3, cheap and fast, and noticeably better than Gemini 4 Argon.

A developer's smell test with zero-library Canvas2D and WGPU particle explosion/morphing demos found good quality, decent speed, and Haiku-like pricing, with a Rust coding loop planned next.

A user finds V4.1 Flash 1017 much better at games and 3D modeling than the previous version, roughly between Opus 5 and 5.5, though it still overthinks with long CoT.