DeepSeek-V4.1-Flash

763.2BChatby DeepSeek Start a discussion

Open-weight vision-language model from DeepSeek, released 2026-09-10, 763.2B total, MIT. The Flash tier of DeepSeek-V4.1, focused on KV-cache compression, roughly 763B total parameters (MoE).

0

Quantized builds

40

Benchmark records

2

Discussions

Quantized builds

Community and official quantized releases; "linked records" counts only benchmarks tied to that exact release.

No quantized builds cataloged yet.

Benchmark data

40 configurations

HardwareFrameworkQuantDecodeVRAM usedSamplesEvidence
NVIDIA RTX PRO 4000 BlackwellSGLangIQ2_XS2.1 tok/s24 GB1L0 Self-reported
NVIDIA RTX PRO 4000 BlackwellSGLangIQ3_S1.8 tok/s24 GB1L0 Self-reported
NVIDIA RTX PRO 4000 BlackwellvLLMIQ2_XS2.2 tok/s24 GB1L0 Self-reported
NVIDIA RTX PRO 4000 BlackwellvLLMIQ3_S1.8 tok/s24 GB1L0 Self-reported
NVIDIA RTX PRO 5000 BlackwellSGLangIQ2_XS4.3 tok/s48 GB1L0 Self-reported
NVIDIA RTX PRO 5000 BlackwellSGLangIQ3_S3.6 tok/s48 GB1L0 Self-reported
NVIDIA RTX PRO 5000 BlackwellvLLMIQ2_XS4.4 tok/s48 GB1L0 Self-reported
NVIDIA DGX SparkSGLangIQ2_XS0.9 tok/s128 GB1L0 Self-reported
NVIDIA DGX SparkSGLangIQ3_S0.7 tok/s128 GB1L0 Self-reported
NVIDIA DGX SparkvLLMIQ2_XS0.9 tok/s128 GB1L0 Self-reported
NVIDIA DGX SparkvLLMIQ3_S0.8 tok/s128 GB1L0 Self-reported
2x NVIDIA RTX 5070 TiSGLangIQ2_XS—32 GB1L0 Self-reported
2x NVIDIA RTX 5070 TiSGLangIQ3_S—32 GB1L0 Self-reported
2x NVIDIA RTX 5070 TivLLMIQ2_XS—32 GB1L0 Self-reported
2x NVIDIA RTX 5070 TivLLMIQ3_S—32 GB1L0 Self-reported
Apple M5 Ultrallama.cppIQ2_XS3.6 tok/s258.1 GB1L0 Self-reported
Apple M5 Ultrallama.cppIQ3_S3 tok/s309.6 GB1L0 Self-reported
Apple M5 Prollama.cppIQ2_XS0.9 tok/s64 GB1L0 Self-reported
Apple M5 Prollama.cppIQ3_S0.8 tok/s64 GB1L0 Self-reported
Apple M4 Prollama.cppIQ2_XS0.8 tok/s64 GB1L0 Self-reported
Apple M4 Prollama.cppIQ3_S0.7 tok/s64 GB1L0 Self-reported
Apple M3 Maxllama.cppIQ2_XS0.9 tok/s36 GB1L0 Self-reported
Apple M3 Maxllama.cppIQ3_S0.8 tok/s36 GB1L0 Self-reported
NVIDIA RTX 3090llama.cppIQ2_XS2.8 tok/s24 GB1L0 Self-reported
NVIDIA RTX 3090llama.cppIQ3_S2.4 tok/s24 GB1L0 Self-reported
NVIDIA RTX 3090vLLMIQ2_XS3.1 tok/s24 GB1L0 Self-reported
NVIDIA RTX 3090vLLMIQ3_S2.6 tok/s24 GB1L0 Self-reported
NVIDIA RTX 3090 Tillama.cppIQ2_XS3.1 tok/s24 GB1L0 Self-reported
NVIDIA RTX 3090 Tillama.cppIQ3_S2.5 tok/s24 GB1L0 Self-reported
NVIDIA RTX 3090 TivLLMIQ2_XS3.3 tok/s24 GB1L0 Self-reported
NVIDIA RTX 3090 TivLLMIQ3_S2.8 tok/s24 GB1L0 Self-reported
NVIDIA RTX 5070 Tillama.cppIQ2_XS2.7 tok/s16 GB1L0 Self-reported
NVIDIA RTX 5070 Tillama.cppIQ3_S2.3 tok/s16 GB1L0 Self-reported
NVIDIA RTX 5070 TivLLMIQ2_XS3 tok/s16 GB1L0 Self-reported
NVIDIA RTX 5070 TivLLMIQ3_S2.5 tok/s16 GB1L0 Self-reported
NVIDIA RTX PRO 5000 BlackwellvLLMIQ3_S3.7 tok/s48 GB1L0 Self-reported
2x NVIDIA RTX 5090vLLMIQ3_S—64 GB1L0 Self-reported
Apple M5 Maxllama.cppIQ3_S1.5 tok/s128 GB1L0 Self-reported
NVIDIA RTX 5090vLLMIQ3_S4.9 tok/s32 GB1L0 Self-reported
NVIDIA RTX 4090vLLMIQ2_XS3.3 tok/s24 GB1L0 Self-reported

40 of these records are not linked to a specific quantized build; the full set is on the search page.

Community discussion

View all discussions