DeepSeek V4.1 Flash hands-on tests and comparisons

A user generated an interactive 3D male anatomy demo page in DeepSeek Harness with a single prompt on V4.1-Flash-1017, calling the result portfolio-quality.

After recreating a painting-crossover video, the user ranks models from GPT 6 sol up to Fable 5.5, placing Gemini 4 Argon second only to Fable 5.5.

A user reports V4.1 Flash now outscores GLM 5.3 and has clearly improved on the earlier version's glitchiness and weak complex-problem handling, praising the new native search/fetch API and pricing.

The author credits shorter outputs and higher agent scores, but notes long agent runs reuse prefixes and still pay prefill on cache misses, recommending exact caching and hard tool budgets.

A short image post arguing DeepSeek-V4.1-Flash delivers a GPT-6-Astra-like experience at DeepSeek's low price, without concrete test data.

Same-prompt test generating a Miku SVG drawing animation; the author finds Gemini 4 Argon mimics human drawing better while V4.1 Flash follows its own interpretation.

Citing official figures, the post says V4.1 Flash 1017 hits 77% on DeepSWE with output at $0.60–$1.20 per million tokens, far below closed flagships, and discusses open-source economics.

Hands-on impression: V4.1 Flash-1017 covers many visual styles with pixel art the strongest, rated about a half-step behind Opus 5.5, making Claude top-ups unnecessary.

AiHubMix ran the same prompt at max reasoning on GPT 6.1 Sol and DeepSeek V4.1 Flash; both took about 25 minutes, costing $4 and $0.16 respectively, with clearly different outputs.

Hands-on: a single 1h50m run in DeepSeek Harness produced 269,823 tokens at ~258 tok/s, with more thinking, longer prompts and clearly better visuals.

Author shares an optimized repo claiming 250-300 tok/s at concurrency 1 and 1000+ tok/s at c8 on four RTX Pro 6000 GPUs, with long context and vision/engram offloading to SSD.

The author deployed DeepSeek V4.1 Flash 1017 with vLLM on 4x DGX Spark, reporting 165.7 tok/s throughput and 98.6 on a self-built 69-scenario tool-use eval, with the recipe published.