DeepSeek V4.1 Flash hands-on tests and comparisons

BugBunny reports DeepSeek V4.1-Flash-1017 found 54 vulnerabilities across 100 PRs on VulnPR-100 for $7.53; the poster likes the cost per finding but wants false-positive rates and unseen-repo results.

A 2.00bpw Strata quant of the 763B model runs on a 24GB RTX 3090 with 64K context, hitting a 15.9 tok/s median and passing all 21 hidden tests in an OpenCode bug-fix task in 44 minutes.

Native llama.cpp support removes the third-party patch; rough memory math puts 1-bit quants on 128GB machines, 3-bit on a 256GB Mac Studio, and 4-bit-plus at just over 300GB.

A user contends DeepSeek 1017's score of 45 undersells it, since it sometimes beats K3.1's 55, and both belong in a higher tier.

On the VulnPR-100 benchmark DeepSeek V4.1 Flash found 54 vulnerable PRs for under $8, while Kimi K3.1 Pro found 47 and cost over $100.

The llama.cpp mainline now runs DeepSeek-V4.1-Flash natively instead of via third-party patches, prompting debate over waiting for quant packages or building from main.

vLLM reports that under DeepSeek's decode configuration, V4.1 Flash performance matches DeepInfer across batch sizes 8 to 256, using DEP32 versus EP64DP64.

SGLang and vLLM each announced same-day that their inference efficiency on DeepSeek V4.1 Flash matches DeepInfer, by integrating Mega-series and DSpark components.

Seven local configurations from a 16GB RTX 5060 Ti to four DGX Sparks with 512GB built interactive browser worlds, with DeepSeek-V4.1-Flash running across two DGX Sparks.

GSQ-RCO quantization with the Strata inference engine runs DeepSeek-V4.1-Flash on 12GB consumer GPUs, hitting 105–117 tok/s on an RTX 5070 with 64GB RAM, with prefill notably faster than decode.

The author applied a heretic ablation to DeepSeek V4.1-Flash-1017, reports no measurable MMLU or HumanEval difference, and is running full benchmarks plus a BFCL comparison against baseline to check tooling ability.

ISTA-DASLab releases a GSQ-RCO quantized GGUF at ~257GB for 2.00bpw, requiring a Strata fork; on an RTX 5090 with 128GB RAM it hits 10–11 tok/s generation and 270–300 tok/s prefill at ~26,950 tokens.