readygpu

Llama 3.1 8B

8B parameters · 128k context · llama3.1 · released 2024-07-18

chat

What to run it on in Canada

Fully in memory at 32k context, priced at retailers today.

Something we price holds it fully in memory, but none reaches 15 tok/s, our bar for recommending it. The table below has the speeds.

All hardware that runs it

Everything we price, best value first. Two cards split the model by layers.

HardwareQuantFitsDecodePricetok/s per 1,000
NVIDIA GeForce RTX 3090Q6_Kfits53 tok/s––
NVIDIA GeForce RTX 5090Q6_Kfits102 tok/s––
Mac Studio M4 Max 128GBQ6_Kfits33 tok/s––
NVIDIA GeForce RTX 4090Q6_Kfits57 tok/s––
NVIDIA GeForce RTX 5070 TiQ6_Kfits51 tok/s––
Mac mini M4 Pro 64GBQ6_Kfits16 tok/s––
NVIDIA GeForce RTX 5060 Ti 16GBQ6_Kfits26 tok/s––
AMD Radeon RX 7900 XTXQ6_Kfits49 tok/s––
Framework Desktop Ryzen AI Max+ 395 128GBQ6_Kfits13 tok/s––
Mac Studio M3 Ultra 512GBQ6_Kfits49 tok/s––
NVIDIA DGX Spark 128GBQ6_Kfits16 tok/s––
NVIDIA GeForce RTX 3060 12GBQ6_Kfits21 tok/s––
AMD Radeon RX 9070 XTQ6_Kfits32 tok/s––
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ6_Kfits13 tok/s––
Mac Studio M3 Ultra 256GBQ6_Kfits49 tok/s––
Mac mini M4 Pro 48GBQ6_Kfits16 tok/s––
NVIDIA GeForce RTX 4060 Ti 16GBQ6_Kfits16 tok/s––
NVIDIA GeForce RTX 5080Q6_Kfits55 tok/s––
Mac Studio M4 Max 64GBQ6_Kfits33 tok/s––
NVIDIA GeForce RTX 4070 Ti SuperQ6_Kfits38 tok/s––
NVIDIA GeForce RTX 5070Q6_Kfits38 tok/s––
AMD Radeon RX 9060 XT 16GBQ6_Kfits16 tok/s––
Mac Studio M3 Ultra 96GBQ6_Kfits49 tok/s––
NVIDIA GeForce RTX 4070 SuperQ6_Kfits29 tok/s––
NVIDIA GeForce RTX 4080 SuperQ6_Kfits42 tok/s––
NVIDIA RTX PRO 6000 BlackwellQ6_Kfits102 tok/s––
AMD Radeon AI PRO R9700Q6_Kfits32 tok/s––
AMD Radeon RX 7800 XTQ6_Kfits32 tok/s––
AMD Radeon RX 7900 XTQ6_Kfits40 tok/s––
ASUS Ascent GX10 128GBQ6_Kfits16 tok/s––
Framework Desktop Ryzen AI Max+ 395 64GBQ6_Kfits13 tok/s––
Intel Arc B580Q6_Kfits21 tok/s––
Mac mini M4 32GBQ6_Kfits7 tok/s––
NVIDIA GeForce RTX 3090 TiQ6_Kfits57 tok/s––
AMD Radeon RX 7600 XTQ6_Kfits15 tok/s––
Mac mini M4 16GBQ5_K_Mfits8 tok/s––
Intel Arc Pro B60Q6_Kfits21 tok/s––
NVIDIA Tesla P40Q6_Kfits20 tok/s––
Intel Arc A770 16GBQ6_Kfits26 tok/s––
NVIDIA GeForce RTX 3080 10GBQ4_K_Moffload48 tok/s––

Memory needed

Weights, KV cache and runtime buffers at each context length (KV cache at fp16).

QuantFile@8k@32k@64k@128k
Q4_K_Mgguf4.6 GB6.3 GB9.5 GB13.7 GB22.0 GB
mlx-4bitmlx4.2 GB6.0 GB9.1 GB13.3 GB21.7 GB
Q5_K_Mgguf5.3 GB7.1 GB10.2 GB14.4 GB22.8 GB
Q6_Kgguf6.1 GB7.9 GB11.0 GB15.2 GB23.6 GB

Run it

Copy a command for the runtime you use.

  • Q4_K_M
    • Ollamaollama run hf.co/unsloth/Llama-3.1-8B-Instruct-GGUF:Q4_K_M
    • LM Studiolms get unsloth/Llama-3.1-8B-Instruct-GGUF@q4_k_m
    • llama.cppllama-server -hf unsloth/Llama-3.1-8B-Instruct-GGUF:Q4_K_M
  • mlx-4bit
    • LM Studiolms get mlx-community/Llama-3.1-8B-Instruct-4bit
    • MLXmlx_lm.server --model mlx-community/Llama-3.1-8B-Instruct-4bit
  • Q5_K_M
    • Ollamaollama run hf.co/unsloth/Llama-3.1-8B-Instruct-GGUF:Q5_K_M
    • LM Studiolms get unsloth/Llama-3.1-8B-Instruct-GGUF@q5_k_m
    • llama.cppllama-server -hf unsloth/Llama-3.1-8B-Instruct-GGUF:Q5_K_M
  • Q6_K
    • Ollamaollama run hf.co/unsloth/Llama-3.1-8B-Instruct-GGUF:Q6_K
    • LM Studiolms get unsloth/Llama-3.1-8B-Instruct-GGUF@q6_k
    • llama.cppllama-server -hf unsloth/Llama-3.1-8B-Instruct-GGUF:Q6_K