DeepSeek V4.1 Flash hands-on tests and comparisons

A used 8x V100 server running a custom Rust + CUDA engine serves eight V4.1 Flash agents at about 40-50 tok/s, roughly $0.20-0.25 per million output tokens.

A demo shows V4.1 Flash rebuilding whole webpages from scratch within seconds of entering a URL, with the site generated live by the model.

lxe's MIT-licensed dsx keeps hot experts in VRAM and cold T16 experts in RAM, hitting 34.9 tok/s decode on two RTX 3090s plus 128GB RAM, with CUDA 13 and ~305GB NVMe required.

A Japanese news roundup cites Wccftech reporting that DeepSeek V4.1-Flash-1017 produced an exploit that escapes Chrome's sandbox, framed as part of escalating AI-driven attacks on the open web.

An analysis argues only 500B of DeepSeek V4.1 Flash's weights need to sit on wafer with 192B of embeddings off-chip, so Cerebras could serve it efficiently on fewer than 10 wafers.

On one sky-island video prompt, DeepSeek V4.1 Flash 1017 cost 3 cents and 4 minutes but produced thick fog, Gemini 4 Argon's 15-cent result ranked second best, and Opus 5.5 took 39 minutes and $4.91 for the best output.

A Japanese user tested DeepSeek V4.1 Flash 1017 as orchestrator with Opus 5.5 as sub-agents: Opus 5.5 sometimes refused mid-task after seeing the whole project, while 1017 refuses far less and can split tasks for Opus to execute.

A new DSH plugin lets the model picker call DeepSeek V4.1 Flash, GLM 5.3 Flash, Kimi K3 and more with no API key or login, with npx setup steps and a GitHub link.

On the 70-question research-level CritPt v2 physics benchmark, GPT-6 Astra scores 83% and Opus 5.5 78%, while DeepSeek V4.1 Flash lands at 50.0%, nearly tied with Grok 4.7's 50.6%.

The author reports reaching 744 tokens per second on DeepSeek V4.1 Flash 1017 with a 97% cache hit rate after inference optimization, without detailing hardware or quantization.

You.com launched an enterprise inference router supporting DeepSeek V4.1 Flash and other open-weight models, with a 99.9% uptime SLA, zero data retention, US-hosted infrastructure, and an OpenAI-compatible API at 0% platform fee.

A user complains that posts boasting huge V4.1 Flash tok/s numbers never state the quantization, making the figures uninformative.