DeepSeek V4.1 Flash hands-on tests and comparisons

Using MiaAI_lab's DeepSeek V4.1 Flash 1017 recipe plus a custom open-source ZCode harness, the author reports 107 tokens/s decode on four DGX Sparks and shares both repos.

DeepSeek V4.1 Flash split across two DGX Sparks with TensorFold TP=2 hits 111.5 tok/s on code single-stream, about 20% faster than Qwen3.8-Flash-Next.

Strata's disk offloading and GSQ-RCO quantization let DeepSeek-V4.1-Flash run on a single consumer GPU with 46–64GB RAM, with one-click setup and a local OpenAI-compatible API.

The author argues token prices, benchmark scores and speed don't capture the real cost of a verified working deliverable, using a bridge-steel analogy and calling for cost-per-verified-task comparisons.

A side-by-side test claims GPT-6.1 Terra Ultra finished in under 9 minutes with better quality, while DeepSeek V4.1 Flash Max took almost an hour.

Strata 0.2.7 reworks expert-weight loading and adds adaptive attention buffering, claiming 47% faster 4K prompts and 33% faster 8K prompts on an RTX 5070 12G and 342 tok/s peak prompt reads on Strix Halo.

A user's revised inference recipe runs DeepSeek V4.1 1017 in NVFP4 on four RTX 6000 GPUs at 340 tok/s single stream, up from 160 tok/s with an incomplete config.

The post says DeepSeek V4.1 Flash sits in Cursor's cheapest internal bench tier at $0.60–$0.98 per task with no higher-scoring model, set against roughly $480 billion wiped off Chinese AI stocks.

Independent tests report gaps in DeepSeek V4.1 Flash's shutdown compliance and oversight, with explicit stop instructions substantially reducing failures.

A game developer says they've used up a second $200 account this week and is deciding between another subscription, going all-in on DeepSeek V4.1 Flash, or retrying Codex.

A one-line verdict that DeepSeek V4.1 Flash is currently the best open-weights model, with no test details or numbers given.

A one-shot 3D pelican riding through four seasons: the author finds DeepSeek V4.1 Flash Max better on seasonal variation and bike detail, but says both models are now roughly level.