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
| Hardware | Framework | Quant | Decode | VRAM used | Samples | Evidence |
|---|---|---|---|---|---|---|
| NVIDIA RTX PRO 4000 Blackwell | SGLang | IQ2_XS | 2.1 tok/s | 24 GB | 1 | L0 Self-reported |
| NVIDIA RTX PRO 4000 Blackwell | SGLang | IQ3_S | 1.8 tok/s | 24 GB | 1 | L0 Self-reported |
| NVIDIA RTX PRO 4000 Blackwell | vLLM | IQ2_XS | 2.2 tok/s | 24 GB | 1 | L0 Self-reported |
| NVIDIA RTX PRO 4000 Blackwell | vLLM | IQ3_S | 1.8 tok/s | 24 GB | 1 | L0 Self-reported |
| NVIDIA RTX PRO 5000 Blackwell | SGLang | IQ2_XS | 4.3 tok/s | 48 GB | 1 | L0 Self-reported |
| NVIDIA RTX PRO 5000 Blackwell | SGLang | IQ3_S | 3.6 tok/s | 48 GB | 1 | L0 Self-reported |
| NVIDIA RTX PRO 5000 Blackwell | vLLM | IQ2_XS | 4.4 tok/s | 48 GB | 1 | L0 Self-reported |
| NVIDIA DGX Spark | SGLang | IQ2_XS | 0.9 tok/s | 128 GB | 1 | L0 Self-reported |
| NVIDIA DGX Spark | SGLang | IQ3_S | 0.7 tok/s | 128 GB | 1 | L0 Self-reported |
| NVIDIA DGX Spark | vLLM | IQ2_XS | 0.9 tok/s | 128 GB | 1 | L0 Self-reported |
| NVIDIA DGX Spark | vLLM | IQ3_S | 0.8 tok/s | 128 GB | 1 | L0 Self-reported |
| 2x NVIDIA RTX 5070 Ti | SGLang | IQ2_XS | — | 32 GB | 1 | L0 Self-reported |
| 2x NVIDIA RTX 5070 Ti | SGLang | IQ3_S | — | 32 GB | 1 | L0 Self-reported |
| 2x NVIDIA RTX 5070 Ti | vLLM | IQ2_XS | — | 32 GB | 1 | L0 Self-reported |
| 2x NVIDIA RTX 5070 Ti | vLLM | IQ3_S | — | 32 GB | 1 | L0 Self-reported |
| Apple M5 Ultra | llama.cpp | IQ2_XS | 3.6 tok/s | 258.1 GB | 1 | L0 Self-reported |
| Apple M5 Ultra | llama.cpp | IQ3_S | 3 tok/s | 309.6 GB | 1 | L0 Self-reported |
| Apple M5 Pro | llama.cpp | IQ2_XS | 0.9 tok/s | 64 GB | 1 | L0 Self-reported |
| Apple M5 Pro | llama.cpp | IQ3_S | 0.8 tok/s | 64 GB | 1 | L0 Self-reported |
| Apple M4 Pro | llama.cpp | IQ2_XS | 0.8 tok/s | 64 GB | 1 | L0 Self-reported |
| Apple M4 Pro | llama.cpp | IQ3_S | 0.7 tok/s | 64 GB | 1 | L0 Self-reported |
| Apple M3 Max | llama.cpp | IQ2_XS | 0.9 tok/s | 36 GB | 1 | L0 Self-reported |
| Apple M3 Max | llama.cpp | IQ3_S | 0.8 tok/s | 36 GB | 1 | L0 Self-reported |
| NVIDIA RTX 3090 | llama.cpp | IQ2_XS | 2.8 tok/s | 24 GB | 1 | L0 Self-reported |
| NVIDIA RTX 3090 | llama.cpp | IQ3_S | 2.4 tok/s | 24 GB | 1 | L0 Self-reported |
| NVIDIA RTX 3090 | vLLM | IQ2_XS | 3.1 tok/s | 24 GB | 1 | L0 Self-reported |
| NVIDIA RTX 3090 | vLLM | IQ3_S | 2.6 tok/s | 24 GB | 1 | L0 Self-reported |
| NVIDIA RTX 3090 Ti | llama.cpp | IQ2_XS | 3.1 tok/s | 24 GB | 1 | L0 Self-reported |
| NVIDIA RTX 3090 Ti | llama.cpp | IQ3_S | 2.5 tok/s | 24 GB | 1 | L0 Self-reported |
| NVIDIA RTX 3090 Ti | vLLM | IQ2_XS | 3.3 tok/s | 24 GB | 1 | L0 Self-reported |
| NVIDIA RTX 3090 Ti | vLLM | IQ3_S | 2.8 tok/s | 24 GB | 1 | L0 Self-reported |
| NVIDIA RTX 5070 Ti | llama.cpp | IQ2_XS | 2.7 tok/s | 16 GB | 1 | L0 Self-reported |
| NVIDIA RTX 5070 Ti | llama.cpp | IQ3_S | 2.3 tok/s | 16 GB | 1 | L0 Self-reported |
| NVIDIA RTX 5070 Ti | vLLM | IQ2_XS | 3 tok/s | 16 GB | 1 | L0 Self-reported |
| NVIDIA RTX 5070 Ti | vLLM | IQ3_S | 2.5 tok/s | 16 GB | 1 | L0 Self-reported |
| NVIDIA RTX PRO 5000 Blackwell | vLLM | IQ3_S | 3.7 tok/s | 48 GB | 1 | L0 Self-reported |
| 2x NVIDIA RTX 5090 | vLLM | IQ3_S | — | 64 GB | 1 | L0 Self-reported |
| Apple M5 Max | llama.cpp | IQ3_S | 1.5 tok/s | 128 GB | 1 | L0 Self-reported |
| NVIDIA RTX 5090 | vLLM | IQ3_S | 4.9 tok/s | 32 GB | 1 | L0 Self-reported |
| NVIDIA RTX 4090 | vLLM | IQ2_XS | 3.3 tok/s | 24 GB | 1 | L0 Self-reported |
40 of these records are not linked to a specific quantized build; the full set is on the search page.