DeepSeek V4.1 Flash hands-on tests and comparisons

A Metal decode-fusion patch raises DeepSeek V4.1 Flash throughput on DwarfStar from 18.5 to 31.3 tok/s, near Unsloth's llama.cpp but still below oMLX's 38.5 tok/s.

DeepSeek's newly open-sourced DeepInfer ranks second in weekly star gains; it runs V4.1 Flash on NVIDIA and Ascend chips but omits parts like the KV cache system.

A blind test of 15 anonymised business emails found DeepSeek V4.1 Flash, Opus 5.5 and the user's own replies all solid, but DeepSeek used a banned greeting twice and invented a promise while Opus broke no rules.

Running stock DeepSeek V4.1 Flash with SGLang on four DGX Sparks in a switchless ring gives 70.0 tok/s on single-stream code and 72.9 tok/s total across 4 streams (26.2 each).

After 10+ 3D tests the author finds Qwen 4 Kalmia decent but still below DeepSeek V4.1 Flash on the same task, roughly MiniMax 3.1 Flash level overall.

A user says a DeepSeek-V4.1-Flash fine-tune tailored to T3 Codes could plausibly replace Opus 5.5 as their daily orchestration driver.

A user reports 500–600 tok/s running DeepSeek V4.1 Flash on an RTX Pro 6000 and thinks the resulting demand is raising prices for that GPU.

Project Maya v1.1.28 maps experts from RAM to let a 24GB RTX 3090 run the 748B model, raising decode speed from 12.9 to 27.8 tok/s with 1M context support.

A user reports over 100 tok/s running DeepSeek V4.1 Flash on a 12GB RTX 5070 using the Strata engine's layer-major chunked reading and GSQ-RCO quantization, with more RAM boosting decode speed.

A user shares Hugging Face links to Maya GGUF quants, including one for DeepSeek V4.1-Flash-1017, for local deployment.

On a 16GB 4080 Super with 128GB RAM at IQ2_XXS, generation rose from ~10 to 16.7–22.7 t/s at 15.4GB VRAM, and official RAM-expert loading adds ~24% agent decode speed.

A user compares DeepSeek V4.1 Flash output in OpenCode and Qoder with the pelican test, posting a video where one pelican comes out fat and the other thin.