readygpu

Phi-4 14B

14.7B parameters · 16k context · mit · released 2024-12-11

chatreasoning

What to run it on in India

Fully in memory at 8k 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 3090Q8_0fits34 tok/s––
NVIDIA GeForce RTX 5090Q8_0fits64 tok/s––
Mac Studio M4 Max 128GBQ8_0fits21 tok/s––
NVIDIA GeForce RTX 4090Q8_0fits36 tok/s––
NVIDIA GeForce RTX 5070 TiQ6_Kfits41 tok/s––
Mac mini M4 Pro 64GBQ8_0fits10 tok/s––
NVIDIA GeForce RTX 5060 Ti 16GBQ6_Kfits20 tok/s––
AMD Radeon RX 7900 XTXQ8_0fits31 tok/s––
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits8 tok/s––
Mac Studio M3 Ultra 512GBQ8_0fits31 tok/s––
NVIDIA DGX Spark 128GBQ8_0fits10 tok/s––
NVIDIA GeForce RTX 3060 12GBQ4_K_Mfits21 tok/s––
AMD Radeon RX 9070 XTQ6_Kfits26 tok/s––
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits8 tok/s––
Mac Studio M3 Ultra 256GBQ8_0fits31 tok/s––
Mac mini M4 Pro 48GBQ8_0fits10 tok/s––
NVIDIA GeForce RTX 4060 Ti 16GBQ6_Kfits13 tok/s––
NVIDIA GeForce RTX 5080Q6_Kfits43 tok/s––
Mac Studio M4 Max 64GBQ8_0fits21 tok/s––
NVIDIA GeForce RTX 4070 Ti SuperQ6_Kfits30 tok/s––
NVIDIA GeForce RTX 5070Q4_K_Mfits39 tok/s––
AMD Radeon RX 9060 XT 16GBQ6_Kfits13 tok/s––
Mac Studio M3 Ultra 96GBQ8_0fits31 tok/s––
NVIDIA GeForce RTX 4070 SuperQ4_K_Mfits30 tok/s––
NVIDIA GeForce RTX 4080 SuperQ6_Kfits33 tok/s––
NVIDIA RTX PRO 6000 BlackwellQ8_0fits64 tok/s––
AMD Radeon AI PRO R9700Q8_0fits20 tok/s––
AMD Radeon RX 7800 XTQ6_Kfits25 tok/s––
AMD Radeon RX 7900 XTQ8_0fits26 tok/s––
ASUS Ascent GX10 128GBQ8_0fits10 tok/s––
Framework Desktop Ryzen AI Max+ 395 64GBQ8_0fits8 tok/s––
Intel Arc B580Q4_K_Mfits22 tok/s––
Mac mini M4 32GBQ8_0fits5 tok/s––
NVIDIA GeForce RTX 3090 TiQ8_0fits36 tok/s––
AMD Radeon RX 7600 XTQ6_Kfits12 tok/s––
Mac mini M4 16GBQ4_K_Mfits7 tok/s––
Intel Arc Pro B60Q8_0fits13 tok/s––
NVIDIA Tesla P40Q8_0fits12 tok/s––
Intel Arc A770 16GBQ6_Kfits20 tok/s––
NVIDIA GeForce RTX 3080 10GBQ4_K_Moffload23 tok/s––

Memory needed

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

QuantFile@8k
Q4_K_Mgguf8.3 GB10.6 GB
mlx-4bitmlx7.7 GB10.0 GB
Q5_K_Mgguf9.7 GB12.0 GB
Q6_Kgguf11.2 GB13.5 GB
mlx-8bitmlx14.5 GB16.8 GB
Q8_0gguf14.5 GB16.8 GB

Run it

Copy a command for the runtime you use.

  • Q4_K_M
    • Ollamaollama run hf.co/unsloth/phi-4-GGUF:Q4_K_M
    • LM Studiolms get unsloth/phi-4-GGUF@q4_k_m
    • llama.cppllama-server -hf unsloth/phi-4-GGUF:Q4_K_M
  • mlx-4bit
    • LM Studiolms get mlx-community/phi-4-4bit
    • MLXmlx_lm.server --model mlx-community/phi-4-4bit
  • Q5_K_M
    • Ollamaollama run hf.co/unsloth/phi-4-GGUF:Q5_K_M
    • LM Studiolms get unsloth/phi-4-GGUF@q5_k_m
    • llama.cppllama-server -hf unsloth/phi-4-GGUF:Q5_K_M
  • Q6_K
    • Ollamaollama run hf.co/unsloth/phi-4-GGUF:Q6_K
    • LM Studiolms get unsloth/phi-4-GGUF@q6_k
    • llama.cppllama-server -hf unsloth/phi-4-GGUF:Q6_K
  • mlx-8bit
    • LM Studiolms get mlx-community/phi-4-8bit
    • MLXmlx_lm.server --model mlx-community/phi-4-8bit
  • Q8_0
    • Ollamaollama run hf.co/unsloth/phi-4-GGUF:Q8_0
    • LM Studiolms get unsloth/phi-4-GGUF@q8_0
    • llama.cppllama-server -hf unsloth/phi-4-GGUF:Q8_0