DeepSeek V4.1 Flash hands-on tests and comparisons

A user posts that among providers tried, Ollama is the fastest for DeepSeek V4.1 Flash, without giving concrete tok/s figures.

A CMU researcher says he bought a B300 for LLM research but is now considering serving DeepSeek V4.1 Flash tokens with DeepSeek's new inference code to fund the rest of his lab.

A user runs EXL3 2.9bpw DeepSeek V4.1 Flash on four DGX Sparks via TensorFold TP4, hitting ~155 tok/s code and ~87 prose at 1M context with 16 streams, and lists four setup gotchas.

Pricing 7 days of real Claude Code logs (5.1B tokens) per model gives $1,965 for Opus 5.5, $983 for Sonnet 5.5, $620 for Gemini 4 Argon, $74 for Haiku 5.5 and $44 for DeepSeek V4.1 Flash, about 45x cheaper, though the author concedes quality differs.

The user says GLM-5.3-Flash feels like a practical, production-ready model while DeepSeek-V4.1-Flash still feels experimental, and reports using Maya with GLM daily.

Testing ISTA-DASLab's IQ2-XXS GGUF on a 32GB Pro4500 and a 16GB 5080 build, the author finds VRAM minus shared weights decides how many experts fit, and more DRAM helps less than more VRAM.

A short scenario-based take: DeepSeek V4.1 Flash is fast and cheap to get started, Kimi K3.1 Pro handles long jobs if you ration quota, and Claude Opus 5.5 is pricey but avoids rework.

A hands-on test has ds-v4.1-flash build a procedural steam-harbour 3D open world with PBR materials, SSAO, SMAA and a day/night cycle all in code, using 152M tokens at 99.8% cache hit.

A user reports running DeepSeek V4.1 Flash for free on OpenCode Zen at roughly 110 tok/s, with a screenshot attached.

On the black-box WebExploitBench offensive-security benchmark, Kimi-K3.1-Pro leads at 46.43%, DeepSeek-V4.1-Flash-1017 is second at 42.86%, and GLM-5.3-Flash third at 35.71%.

On Mazebench, Kimi K3.1 Pro is the first open-weights model with a nontrivial score at 2%, while DeepSeek V4.1 Flash follows at 1%.

A user runs V4.1 Flash on an ordinary 64GB-RAM dev desktop using Strata plus gsq-rco quants, arguing that shrinking open weights make per-token mid-tier pricing harder to justify.