DeepSeek V4.1 Flash hands-on tests and comparisons

An EXL3 3.5bpw uncensored quant of DeepSeek-V4.1-Flash-1017 runs at about 50 tok/s decode on 7x3090 plus 256GB RAM using the Project Maya engine, with roughly 260GB of weights.

The poster notes DeepSeek is the only major lab publishing per-GPU-hour costs, citing $38/GPU-hour for V4.1 Flash while Anthropic and OpenAI still lose money.

User jvr0x says that after getting a third Spark they expected to run something like GLM 5.3, but DeepSeek V4.1 Flash reaches the same intelligence level on just two Sparks.

A cost analysis after DeepInfer's open-sourcing compares frontier and open-model token pricing, putting DeepSeek V4.1 Flash at $0.0149 per 1M tokens off-peak versus ~$0.0026 B300 rental cost.

The same pelican-on-a-bicycle prompt is run through DeepSeek V4.1 Flash 1017, GPT-6.1 Sol, Gemini 4 Argon and Claude Opus 5.5, with the author saying the new model's details keep up.

Same prompt, one try each: a 2D Halloween animation of a pigeon on a unicycle in a pumpkin hat; the author calls DeepSeek clean and atmospheric, MiniMax the most detailed, MiMo the weakest, and picks MiniMax overall.

V4.1 Flash posts 45 at $0.26 per task versus Grok 4.7's 46 at $2.73, Qwen3.8 Max's 45 at $5.41 and Kimi K3's 44 at $2.00, making it the price leader among open-weights models.

The author reports the Maya-GGUF quant of DeepSeek V4.1 Flash 1017 has reached 23k Hugging Face downloads and keeps growing.

Fireworks says DeepSeek V4.1 Flash 1017 reaches 1,521+ tok/s on its Ultrafast tier, shipping Tetris in 11 seconds and running about 7x faster than Claude Opus 5.5.

A kernel optimization on four CMP 170HX cards speeds up DeepSeek V4.1-Flash-1017 decode by 2–3% over the author's previous best; against 4× DGX Spark it is 23% faster at 8-user decode but loses on cold prefill.

A curated list of 60 open-weight coding models and runtimes notes that this week's DeepSeek V4.1-Flash, Reflection Beam and LFM2.5-24B have no dedicated GitHub repos, so the runtime tools are what matter.

On offensive-security benchmarks DeepSeek V4.1 cost 0.1× Opus 5 at higher effort while reaching 62.9% recall versus 64.8%, which the poster calls the best cost-to-findings ratio they have seen.