DeepSeek V4.1 Flash hands-on tests and comparisons

Wally brings DeepSeek V4.1 Flash Uncensored to its platform with open-source security harnesses and hosted inference for authorized defensive security work.

Switching four local models in Strata on the same task, the tester found DeepSeek V4.1 Flash strongest on long text and instructions but slow, while Qwen3.8 Coder was fastest at code but broke on Japanese.

By decoding busy experts' weights once to FP16 per chunk and letting cuBLAS do the math, the author raised prefill on a single RTX 5090 from 1,400 to 1,950 t/s (+39%), shipped as PR #4366 to the Strata repo.

Input prices range from $0.00848 to $0.60 per million tokens and output from $0.18 to $2.40, with cheap tiers flagged as low precision or idle compute.

The author explains why output tokens cost 3–5x input using OpenRouter prices, listing DeepSeek V4.1 Flash at $0.14/$0.42 versus Opus 5.5 at $4/$20.

A user shows local DeepSeek V4.1 Flash generating charts and diagrams by writing Python scripts that render PNGs, then using vision to review and fix the output, all on-device.

The new coding benchmark from 20+ maintainers leaves requirements implicit and zeroes hard-rule violations; DeepSeek V4.1 Flash scores 19.3% against Opus 5.5's 45.5%.

A user shows DeepSeek-v4.1-flash-1017 running locally on their two DGX Sparks inside the DeepSeek harness.

A user says V4.1 Flash 10-17 performs around Opus 4.6–4.8 level but is very slow and makes worse guesses and theories than Claude.

A user got DeepSeek V4.1 Flash Maya XS running on an AMD Ryzen AI Max+ 395 DIVE CX box at about 15 tok/s, with setup and tuning done by GPT-6 Astra, and notes remaining issues.

Using OpenCode with DeepSeek V4.1 Flash, the author spent under $2 and about two hours on modeling, texturing and Blender rendering, then had it generate a Godot project from the same asset.

The author compared eight models on small customer-message handling: GLM 5.3 Flash usable but slow, Haiku not smart enough, Sonnet/Opus too expensive, GPT 6.1 Terra weak, Grok 4.7 good but slow; the two Chinese models won on very low price and fast replies, while Opus 5.5 remains the pick for code.