DeepSeek V4.1 Flash hands-on tests and comparisons

Running the 763B-parameter DeepSeek V4.1 Flash on a 24GB L4 GPU via Strata yields 2.34 tok/s — slow, but notable for fitting a model that size on one card.

A developer says three days of round-the-clock bug hunting would have cost about $800 in DeepSeek V4.1 Flash 1017 API credits (over $6,000 on Opus), but ran locally on a dual-GB10 host at concurrency 4 and 512k context.

A user reports that running DeepSeek V4.1 Flash-1017 locally on four Sparks delivers both speed and intelligence; Opus 5.5 is still better, but fully local work is now viable, though it saves neither money nor time.

A user shares a TP=2 DeepSeek V4.1 Flash quant for two DGX Sparks: EXL3 at 2.80 BPW, 1M context per chat, and 93/100 on their own 353-question MMLU-Pro STEM set.

A user says the iOS OpenBot app can run DeepSeek V4.1 Flash 1017 and Qwen Flash locally, with a video demo.

The author argues DeepInfer mainly benefits hosting vendors: it only supports datacenter cards like B200 and Ascend 950 plus offline DeepSeek-V4.1-Flash inference, while individual devs gain indirectly via lower costs and more providers.

Decode throughput test: SGLang baseline 71 tps, tuned 103 tps, and TensorFold tuned 110 tps.

Under TP4 ring, 256K input and thinking on, the knapcio SGLang recipe matched MiaAI's SGLang decode and was about 23% faster in prefill, while the TensorFold build lagged.

The author tries a DeepSeek V4.1 Flash 1017 IQ2_XXS-Uncensored quant on Strata, swapping the SSD-resident Engram part from 4-bit to original MXFP8 weights using only the weights already on hand.

Same prompt to rebuild the Apple website: LFM3 ran locally for $0 while DeepSeek V4.1 Flash cost $0.07, and the author calls LFM3's output insane.

A user running DeepSeek 1017 locally reports peak single-stream decoding of 90 tok/s and says the local model beats the online version without quality degradation.

A user reports running a 748B DeepSeek V4.1 Flash build locally on stock llama.cpp and links GSQ-RCO upstream GGUF quantized weights on Hugging Face.