DeepSeek V4.1 Flash hands-on tests and comparisons

A single-attempt rerun of the same prompt on two DGX Sparks via TensorFold Zig finished in 74.8 minutes, producing 88,201 visible output tokens out of 247,134 generated including reasoning, with a repro repo.

After DeepSeek open-sourced DeepInfer, the engine serving V4.1 Flash, a team got vLLM to match it on the official decode config using half the GPUs, then scaled to 954K tokens/s per GPU across a GB300 NVL72 rack.

A Polish blogger says the open-weight V4.1 Flash 1017 tops open-source models at 57% on Artificial Analysis' cyber index with zero refusals, at about 1.65 PLN per task, and criticizes Anthropic for limiting Mythos to roughly 50 firms.

On the same prompt the author ranks Opus 5.5 ≈ GPT 6.1 > Kimi K3.1 Pro > DeepSeek V4.1 > MiniMax, calling Kimi's result the best overall.

A user shares a VulcanBench comparison of Opus 5.5, GPT-6.1 Sol and DeepSeek V4.1 Flash, noting DeepSeek's blue bars are almost nonexistent.

A walkthrough for adding a plugin to DeepSeek Harness so DeepSeek V4.1 Flash, GLM 5.3 Flash and Kimi K3 appear in the model picker for free, with no login or API key, subject to region and lineup.

The poster burned 860M V4.1 Flash tokens in 3.5 hours, 154M of them on a custom IDE, and used the model inside a Windows VM to take apart Diablo IV.

Using P/D disaggregation and Wide-EP on a GB300 NVL72 rack, vLLM serves V4.1-Flash on SemiAnalysis AgentX at 954K tokens/s per GPU with 196 tokens/s P90 per user, 7.9x its single-node result.

Strata's terminal build bundles V4.1 Flash and serves a local OpenAI/Anthropic API; on a 12GB RTX 5070 with 64GB RAM, Q2_0 hits 94 tok/s and IQ3_S 53 tok/s, with prompt intake around 2,650 tok/s.

A solo non-CS developer splits a Ghostty terminal into four panes via Herdr and OMP: Opus 5.5 dispatches, V4.1 Flash 1017 Max writes code, Sol and Astra escalate and review; cost savings are not yet measured.

Qoder users can use DeepSeek V4.1 Flash 1017 free through November 1, with Qwen3.8-Flash free indefinitely and 100 daily credits for GLM 5.3, Kimi K3.1 Pro and others.

From limited input the author finds V4.1 Flash the best conversational model, agreeing that GPT-6.1 Sol often sounds forced and trained to be relatable.