DeepSeek V4.1 Flash hands-on tests and comparisons

Korean analysis of DeepSeek's own comparison table: 77.0 on DeepSWE vs Opus 5's 74.0, but 42.6 on HLE vs 56.3; API at $0.15/$0.60 per million tokens, roughly 1/16–1/32 of GPT-5.6 Sol.

Same code-only 3D pelican riding a bike prompt: Gemini handles lighting and kinematics better, while DeepSeek's geometry is rough with disjointed motion and clipping.

A user asks whether DeepSeek V4.1 Flash 1017 was distilled from GPT-6 Astra, saying it feels increasingly similar, with no evidence given.

The author works through deployment costs: 763B parameters and roughly 510GB of weights mean a 4060, 4090 or even 5090 cannot hold the full model, enterprise multi-GPU setups may cost millions, and individuals are better off with the API.

A 24-hour AI news digest noting V4.1 Flash 1017 ships MIT open weights, scores 77.0 on DeepSWE and 42.3 on Terminal-Bench 4.0, keeps API pricing, and adds Search/Fetch APIs.

The author is tuning six AI compute boxes in parallel: two for DeepSeek V4.1 Flash 1017, two for GLM 5.3 Flash FP4 local, and two for Qwen3.8 Flash Next FP4 QAD quantization.

Official info: live on web, app and API under the same deepseek-flash name and pricing, with better benchmarks, shorter outputs and open weights on HuggingFace and ModelScope.

The Shiguredo fork of shiguredo/deepseek-harness now supports DeepSeek V4.1 Flash 1017 on Ollama Cloud, with a gist listing the changes.

Billed as a minor update, it improves several benchmarks including Terminal-Bench at unchanged API pricing and opens a self-built Search API with hundreds of millions of daily queries.

Chinese write-up: the new V4.1-Flash mode is faster and cheaper, using self-feedback RL for one-sentence project building, with a filtered web search API at ¥5 per 1,000 calls, free in Harness.

Chat app Talelynx says after DeepSeek's update, V4.1 Flash (+Thinking) loses character settings in long sessions; it added config tweaks and recommends starting a new session.

Local test: Pagoda Garden never appeared even at 110K tokens, and speed dropped to 5.7 tok/s on an M3 Mac Studio, so it merely runs.