readygpu

DeepSeek R1 Distill Qwen 14B

14.8B parameters · 128k context · MIT License · released 2025-01-20

reasoningchatcoding

In short

DeepSeek R1 Distill Qwen 14B is an open model from DeepSeek with 14.8B parameters. At mlx-3bit, DeepSeek R1 Distill Qwen 14B needs about 8 GB of memory at 8k context and 13 GB at 32k, counting the weights, the KV cache and runtime buffers. The cheapest hardware in the US that runs DeepSeek R1 Distill Qwen 14B fully in memory at 32k context is the NVIDIA GeForce RTX 5060 Ti 16GB, $618 at eBay, at about 18 tok/s (estimated).

Updated . Speeds are estimates; how they are worked out.

What to run it on in the US

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

All hardware that runs it

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

HardwareQuantFitsDecodePricetok/s per 1,000
NVIDIA Tesla P40Q8_0fits10 tok/s$34528.1
NVIDIA GeForce RTX 5060 Ti 16GBQ4_K_Mfits18 tok/s$61829.1
Intel Arc Pro B60Q8_0fits10 tok/s$65015.8
NVIDIA GeForce RTX 4060 Ti 16GBQ4_K_Mfits12 tok/s$77015.1
AMD Radeon RX 7900 XTQ6_Kfits24 tok/s$80029.6
NVIDIA GeForce RTX 4070 Ti SuperQ4_K_Mfits27 tok/s$98427.5
NVIDIA GeForce RTX 5070 TiQ4_K_Mfits36 tok/s$99536.2
NVIDIA GeForce RTX 4080 SuperQ4_K_Mfits30 tok/s$1,19924.7
Mac mini M4 32GBQ6_Kfits4 tok/s$1,4203.0
AMD Radeon RX 7900 XTXQ8_0fits24 tok/s$1,53315.5
NVIDIA GeForce RTX 3090Q8_0fits26 tok/s$1,60016.4
NVIDIA GeForce RTX 5080Q4_K_Mfits39 tok/s$1,60024.1
AMD Radeon AI PRO R9700Q8_0fits16 tok/s$1,7009.4
NVIDIA GeForce RTX 3090 TiQ8_0fits28 tok/s$2,06013.7
NVIDIA GeForce RTX 4090Q8_0fits28 tok/s$2,07313.6
Mac mini M4 Pro 48GBQ8_0fits8 tok/s$2,1493.7
Mac mini M4 Pro 64GBQ8_0fits8 tok/s$2,9002.8
Mac Studio M4 Max 64GBQ8_0fits16 tok/s$3,4004.7
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits6 tok/s$3,4491.9
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits6 tok/s$3,6501.8
Mac Studio M3 Ultra 96GBQ8_0fits24 tok/s$4,3995.5
NVIDIA GeForce RTX 5090Q8_0fits50 tok/s$4,92110.2
Mac Studio M4 Max 128GBQ8_0fits16 tok/s$4,9993.2
ASUS Ascent GX10 128GBQ8_0fits8 tok/s$6,1001.2
NVIDIA DGX Spark 128GBQ8_0fits8 tok/s$7,3991.0
NVIDIA RTX PRO 6000 BlackwellQ8_0fits50 tok/s$7,9976.3
Mac Studio M3 Ultra 256GBQ8_0fits24 tok/s$8,4992.8
Mac Studio M3 Ultra 512GBQ8_0fits24 tok/s$23,0001.0
Framework Desktop Ryzen AI Max+ 395 64GBQ8_0fits6 tok/s––
Intel Arc B580Q4_K_Moffload8 tok/s$34622.0
NVIDIA GeForce RTX 3060 12GBQ4_K_Moffload8 tok/s$36021.7
Intel Arc A770 16GBQ4_K_Moffload17 tok/s$40043.3
NVIDIA GeForce RTX 3080 10GBQ4_K_Moffload7 tok/s$46915.6
AMD Radeon RX 7600 XTQ4_K_Moffload10 tok/s$48020.8
AMD Radeon RX 9060 XT 16GBQ4_K_Moffload11 tok/s$52021.3
AMD Radeon RX 7800 XTQ4_K_Moffload21 tok/s$55038.2
NVIDIA GeForce RTX 4070 SuperQ4_K_Moffload9 tok/s$69912.6
NVIDIA GeForce RTX 5070Q4_K_Moffload10 tok/s$75012.8
AMD Radeon RX 9070 XTQ4_K_Moffload22 tok/s$81526.4

1 other card or machine we price cannot hold it at this context.

Memory needed

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

QuantFile@8k@32k@64k@128k
mlx-3bitmlx6.0 GB8.3 GB12.9 GB19.1 GB31.5 GB
mlx-4bitmlx7.7 GB10.0 GB14.6 GB20.8 GB33.2 GB
Q4_K_Mgguf8.4 GB10.6 GB15.3 GB21.4 GB33.8 GB
Q5_K_Mgguf9.8 GB12.0 GB16.7 GB22.9 GB35.2 GB
mlx-6bitmlx11.2 GB13.4 GB18.1 GB24.3 GB36.6 GB
Q6_Kgguf11.3 GB13.5 GB18.2 GB24.4 GB36.7 GB
mlx-8bitmlx14.6 GB16.9 GB21.5 GB27.7 GB40.1 GB
Q8_0gguf14.6 GB16.9 GB21.5 GB27.7 GB40.1 GB

Run it

Copy a command for the runtime you use.

  • mlx-3bit
    • LM Studiolms get mlx-community/DeepSeek-R1-Distill-Qwen-14B-3bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Qwen-14B-3bit
  • mlx-4bit
    • LM Studiolms get mlx-community/DeepSeek-R1-Distill-Qwen-14B-4bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Qwen-14B-4bit
  • Q4_K_M
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q4_K_M
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF@q4_k_m
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q4_K_M
  • Q5_K_M
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q5_K_M
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF@q5_k_m
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q5_K_M
  • mlx-6bit
    • LM Studiolms get mlx-community/DeepSeek-R1-Distill-Qwen-14B-6bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Qwen-14B-6bit
  • Q6_K
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q6_K
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF@q6_k
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q6_K
  • mlx-8bit
    • LM Studiolms get mlx-community/DeepSeek-R1-Distill-Qwen-14B-8bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Qwen-14B-8bit
  • Q8_0
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q8_0
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF@q8_0
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q8_0

Notes

Distilled from DeepSeek-R1 onto Qwen2.5-14B (originally Apache 2.0); card says to avoid system prompts and use temperature 0.6. Card recommends forcing output to start with "<think>\n".

Questions

How much VRAM does DeepSeek R1 Distill Qwen 14B need?
At mlx-3bit, DeepSeek R1 Distill Qwen 14B needs about 8 GB of memory at 8k context and 13 GB at 32k, counting the weights, the KV cache and runtime buffers.
What is the best GPU for running DeepSeek R1 Distill Qwen 14B locally?
The cheapest hardware in the US that runs DeepSeek R1 Distill Qwen 14B fully in memory at 32k context is the NVIDIA GeForce RTX 5060 Ti 16GB, $618 at eBay, at about 18 tok/s (estimated). The best value is the NVIDIA GeForce RTX 5070 Ti (about 36 tok/s for $995). The fastest we price is the NVIDIA GeForce RTX 5090, at about 50 tok/s.
Can I run DeepSeek R1 Distill Qwen 14B on a 16 GB GPU?
Yes: at Q4_K_M it fits fully in 16 GB at 32k context. On the NVIDIA GeForce RTX 5080 it runs at about 39 tok/s (estimated).
Can I run DeepSeek R1 Distill Qwen 14B on a 24 GB GPU?
Yes: at Q8_0 it fits fully in 24 GB at 32k context. On the NVIDIA GeForce RTX 3090 Ti it runs at about 28 tok/s (estimated).
How do I run DeepSeek R1 Distill Qwen 14B locally?
With Ollama: ollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q4_K_M (the Q4_K_M build). LM Studio and llama.cpp commands are on this page.
What licence is DeepSeek R1 Distill Qwen 14B released under?
MIT License, which allows commercial use.