What Models?

VRAM needed per quantization

Weights plus fp16 KV cache. Leave ~1 GB headroom for the inference engine.

Cheapest GPUs that can run Codestral 22B

Lowest launch MSRP among GPUs that run its smallest quantization (Q4_K_M) well, with at least 16K context.

• Radeon RX 7900 XT 20 GB · ~56 tok/s at Q4_K_M · Launch MSRP $899
  Check price on Amazon
• Radeon RX 7900 XTX 24 GB · ~67 tok/s at Q4_K_M · Launch MSRP $999
  Check price on Amazon
• RTX 3090 24 GB · ~65 tok/s at Q4_K_M · Launch MSRP $1,499
  Check price on Amazon

As an Amazon Associate, I earn from qualifying purchases.

GPUs that can run Codestral 22B

Best quantization that fits on each GPU, with the context and speed it gets. For chips with several memory sizes, the smallest that works is shown.

NVIDIA

• RTX 4060 Ti 16GB 16 GB Q4_K_M 12K context · ~20 tok/s tight fit
• RTX 4070 Ti SUPER 16 GB Q4_K_M 12K context · ~47 tok/s tight fit
• RTX 4080 16 GB Q4_K_M 12K context · ~50 tok/s tight fit
• RTX 4080 SUPER 16 GB Q4_K_M 12K context · ~51 tok/s tight fit
• RTX 5060 Ti 16GB 16 GB Q4_K_M 12K context · ~31 tok/s tight fit
• RTX 5070 Ti 16 GB Q4_K_M 12K context · ~63 tok/s tight fit
• RTX 5080 16 GB Q4_K_M 12K context · ~67 tok/s tight fit
• RTX 2000 Ada 16 GB Q4_K_M 12K context · ~16 tok/s tight fit
• RTX A4000 16 GB Q4_K_M 12K context · ~31 tok/s tight fit
• RTX PRO 2000 Blackwell 16 GB Q4_K_M 12K context · ~20 tok/s tight fit
• Tesla T4 16 GB Q4_K_M 12K context · ~22 tok/s tight fit
• Tesla V100 16GB 16 GB Q4_K_M 12K context · ~63 tok/s tight fit
• RTX 4000 Ada 20 GB Q4_K_M 31K context · ~25 tok/s
• RTX 3090 24 GB Q4_K_M 32K context · ~65 tok/s
• RTX 3090 Ti 24 GB Q4_K_M 32K context · ~70 tok/s
• RTX 4090 24 GB Q4_K_M 32K context · ~70 tok/s
• RTX 4500 Ada 24 GB Q4_K_M 32K context · ~30 tok/s
• RTX A5000 24 GB Q4_K_M 32K context · ~54 tok/s
• RTX PRO 4000 Blackwell 24 GB Q4_K_M 32K context · ~47 tok/s
• RTX Spark 24 GB Q4_K_M 32K context · ~21 tok/s
• Tesla P40 24 GB Q4_K_M 32K context · ~24 tok/s
• RTX 5090 32 GB Q8_0 32K context · ~73 tok/s
• RTX 5000 Ada 32 GB Q8_0 32K context · ~23 tok/s
• RTX PRO 4500 Blackwell 32 GB Q8_0 32K context · ~36 tok/s
• Tesla V100 32GB 32 GB Q8_0 32K context · ~37 tok/s
• L40S 48 GB Q8_0 32K context · ~35 tok/s
• RTX 6000 Ada 48 GB Q8_0 32K context · ~39 tok/s
• RTX A6000 48 GB Q8_0 32K context · ~31 tok/s
• RTX PRO 5000 Blackwell 48 GB Q8_0 32K context · ~55 tok/s
• RTX PRO 5000 Blackwell 72GB 72 GB Q8_0 32K context · ~55 tok/s
• RTX PRO 5500 Blackwell 84 GB Q8_0 32K context · ~57 tok/s
• RTX PRO 6000 Blackwell 96 GB Q8_0 32K context · ~73 tok/s
• DGX Spark (GB 10) 128 GB Q8_0 32K context · ~11 tok/s

AMD

• AMD BC-250 16 GB Q4_K_M 12K context · ~31 tok/s tight fit
• Radeon RX 6800 16 GB Q4_K_M 12K context · ~36 tok/s tight fit
• Radeon RX 6800 XT 16 GB Q4_K_M 12K context · ~36 tok/s tight fit
• Radeon RX 6900 XT 16 GB Q4_K_M 12K context · ~36 tok/s tight fit
• Radeon RX 6950 XT 16 GB Q4_K_M 12K context · ~40 tok/s tight fit
• Radeon RX 7600 XT 16 GB Q4_K_M 12K context · ~20 tok/s tight fit
• Radeon RX 7800 XT 16 GB Q4_K_M 12K context · ~44 tok/s tight fit
• Radeon RX 7900 GRE 16 GB Q4_K_M 12K context · ~40 tok/s tight fit
• Radeon RX 9060 XT 16GB 16 GB Q4_K_M 12K context · ~22 tok/s tight fit
• Radeon RX 9070 16 GB Q4_K_M 12K context · ~45 tok/s tight fit
• Radeon RX 9070 XT 16 GB Q4_K_M 12K context · ~45 tok/s tight fit
• Radeon RX 7900 XT 20 GB Q4_K_M 31K context · ~56 tok/s
• Radeon RX 7900 XTX 24 GB Q4_K_M 32K context · ~67 tok/s
• Radeon AI PRO R9700 32 GB Q8_0 32K context · ~26 tok/s
• Ryzen AI Max (Strix Halo) 32 GB Q8_0 32K context · ~10 tok/s
• Ryzen AI Max 400 (Gorgon Halo) 64 GB Q8_0 32K context · ~11 tok/s

Intel

• Arc A770 16GB 16 GB Q4_K_M 12K context · ~39 tok/s tight fit
• Arc Pro B65 32 GB Q8_0 32K context · ~25 tok/s
• Arc Pro B70 32 GB Q8_0 32K context · ~25 tok/s

Apple

• M1 16 GB Q4_K_M 12K context · ~5 tok/s tight fit
• M3 Pro 18 GB Q4_K_M 21K context · ~10 tok/s
• M2 24 GB Q4_K_M 32K context · ~7 tok/s
• M3 24 GB Q4_K_M 32K context · ~7 tok/s
• M4 24 GB Q4_K_M 32K context · ~8 tok/s
• M4 Pro 24 GB Q4_K_M 32K context · ~19 tok/s
• M5 24 GB Q4_K_M 32K context · ~11 tok/s
• M5 Pro 24 GB Q4_K_M 32K context · ~21 tok/s
• M6 24 GB Q4_K_M 32K context · ~12 tok/s
• M1 Max 32 GB Q8_0 32K context · ~16 tok/s
• M1 Pro 32 GB Q8_0 32K context · ~8 tok/s
• M2 Max 32 GB Q8_0 32K context · ~16 tok/s
• M2 Pro 32 GB Q8_0 32K context · ~8 tok/s
• M3 Max 36 GB Q8_0 32K context · ~12 tok/s
• M4 Max 36 GB Q8_0 32K context · ~17 tok/s
• M5 Max 36 GB Q8_0 32K context · ~19 tok/s
• M1 Ultra 64 GB Q8_0 32K context · ~32 tok/s
• M2 Ultra 64 GB Q8_0 32K context · ~32 tok/s
• M3 Ultra 96 GB Q8_0 32K context · ~33 tok/s
• M5 Ultra 96 GB Q8_0 32K context · ~49 tok/s