DeepSeek V4.1 Flash hands-on tests and comparisons

A TensorFold recipe runs an EXL3 2.5bpw quant of DeepSeek V4.1 Flash with TP2 on two DGX Sparks, reaching 70.6 tok/s decode and 17.9s to first token on a 32K prompt, with MMLU-Pro 62.0%, LiveCodeBench 60.0% and Tool Eval 93/100.

A same-prompt, same-task video comparison of DeepSeek V4.1 Flash 1017, Minimax M3.1 Flash and Kimi K3.1 Pro, with the verdict left to viewers.

Running the 256-expert DeepSeek V4.1 Flash on one RTX 5090 plus 121GB RAM reaches double-digit tok/s; benchmarking two engines shows RAM placement, not the engine, is the deciding factor.

A user testing Kimi K3.1 Pro hit its usage limit after about 20 minutes and concluded DeepSeek V4.1 Flash 1017 is the better value at that price.

On Artificial Analysis' public cyber index, Kimi K3.1 Pro scores 64, DeepSeek V4.1 Flash 1017 scores 57 and Grok 4.7 scores 56; the post notes DeepSeek had zero safety refusals and that DHS ordered it off civilian agency systems.

Graded on whether a maintainer would merge the change, Codex with DeepSeek V4.1 Flash scores 52.6% versus 56.0% for Claude Code with Opus 5.5, and a quantized build runs on a 64GB home PC with a 12GB GPU via Strata.

V4.1 Flash charges $0.15/$0.60 per million tokens off-peak and double at peak, so even its priciest hour is about 17x cheaper on output than Opus 5.5's $4/$20.

DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF is a mixed-precision MoE quant aimed at local inference.

The poster asks the community for impressions of DeepSeek V4.1 Flash, specifically whether it really beats GPT-5.6 Sol in coding and whether the reasoning modes differ; no test results are included.

gatitolabs congratulates Kimi on K3.1 Pro and posts ModelMeter benchmark charts comparing it with DeepSeek V4.1 Flash.

A user reports OpenCode's daily 200 free requests no longer cover even a small feature workflow with DeepSeek V4.1 Flash, and the PPPoE IP-rotation workaround stopped working.

Developer AkitaOnRails says DeepSeek V4.1 Flash now comes close to Sonnet 5 or Luna and is more than enough for daily tasks, so he will use it more, linking his open-source LLM benchmark writeup.