readygpu

DeepSeek R1 Distill Qwen 1.5B

1.8B parameters · 128k context · MIT License · released 2025-01-20

reasoningchat

In short

DeepSeek R1 Distill Qwen 1.5B is an open model from DeepSeek with 1.8B parameters. At mlx-3bit, DeepSeek R1 Distill Qwen 1.5B needs about 2 GB of memory at 8k context and 6 GB at 128k, counting the weights, the KV cache and runtime buffers. The cheapest hardware in the US that runs DeepSeek R1 Distill Qwen 1.5B fully in memory at 128k context is the NVIDIA Tesla P40, $345 at Amazon, at about 38 tok/s (estimated).

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

What to run it on in the US

Fully in memory at 128k 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_0fits38 tok/s$345110.1
Intel Arc B580Q8_0fits40 tok/s$346116.5
NVIDIA GeForce RTX 3060 12GBQ8_0fits40 tok/s$360109.7
Intel Arc A770 16GBQ8_0fits50 tok/s$400123.8
NVIDIA GeForce RTX 3080 10GBQ8_0fits83 tok/s$469177.7
AMD Radeon RX 7600 XTQ8_0fits28 tok/s$48058.3
AMD Radeon RX 9060 XT 16GBQ8_0fits31 tok/s$52059.8
AMD Radeon RX 7800 XTQ8_0fits61 tok/s$550110.4
NVIDIA GeForce RTX 5060 Ti 16GBQ8_0fits49 tok/s$61879.5
Intel Arc Pro B60Q8_0fits40 tok/s$65062.0
NVIDIA GeForce RTX 4070 SuperQ8_0fits55 tok/s$69979.1
Mac mini M4 16GBQ8_0fits14 tok/s$74918.4
NVIDIA GeForce RTX 5070Q8_0fits74 tok/s$75098.3
NVIDIA GeForce RTX 4060 Ti 16GBQ8_0fits32 tok/s$77041.0
AMD Radeon RX 7900 XTQ8_0fits78 tok/s$80097.3
AMD Radeon RX 9070 XTQ8_0fits62 tok/s$81576.4
NVIDIA GeForce RTX 4070 Ti SuperQ8_0fits74 tok/s$98474.9
NVIDIA GeForce RTX 5070 TiQ8_0fits98 tok/s$99598.8
NVIDIA GeForce RTX 4080 SuperQ8_0fits81 tok/s$1,19967.3
Mac mini M4 32GBQ8_0fits14 tok/s$1,4209.7
AMD Radeon RX 7900 XTXQ8_0fits93 tok/s$1,53360.9
NVIDIA GeForce RTX 3090Q8_0fits103 tok/s$1,60064.2
NVIDIA GeForce RTX 5080Q8_0fits105 tok/s$1,60065.8
AMD Radeon AI PRO R9700Q8_0fits62 tok/s$1,70036.6
NVIDIA GeForce RTX 3090 TiQ8_0fits111 tok/s$2,06053.7
NVIDIA GeForce RTX 4090Q8_0fits111 tok/s$2,07353.3
Mac mini M4 Pro 48GBQ8_0fits31 tok/s$2,14914.6
Mac mini M4 Pro 64GBQ8_0fits31 tok/s$2,90010.8
Mac Studio M4 Max 64GBQ8_0fits63 tok/s$3,40018.5
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits25 tok/s$3,4497.2
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits25 tok/s$3,6506.8
Mac Studio M3 Ultra 96GBQ8_0fits94 tok/s$4,39921.4
NVIDIA GeForce RTX 5090Q8_0fits197 tok/s$4,92140.0
Mac Studio M4 Max 128GBQ8_0fits63 tok/s$4,99912.6
ASUS Ascent GX10 128GBQ8_0fits30 tok/s$6,1004.9
NVIDIA DGX Spark 128GBQ8_0fits30 tok/s$7,3994.0
NVIDIA RTX PRO 6000 BlackwellQ8_0fits197 tok/s$7,99724.6
Mac Studio M3 Ultra 256GBQ8_0fits94 tok/s$8,49911.1
Mac Studio M3 Ultra 512GBQ8_0fits94 tok/s$23,0004.1
Framework Desktop Ryzen AI Max+ 395 64GBQ8_0fits25 tok/s––

Memory needed

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

QuantFile@8k@32k@64k@128k
mlx-3bitmlx0.7 GB1.7 GB2.5 GB3.5 GB5.7 GB
mlx-4bitmlx0.9 GB1.9 GB2.7 GB3.8 GB5.9 GB
Q4_K_Mgguf1.0 GB2.0 GB2.8 GB3.9 GB6.0 GB
Q5_K_Mgguf1.2 GB2.2 GB3.0 GB4.0 GB6.1 GB
mlx-6bitmlx1.3 GB2.3 GB3.1 GB4.2 GB6.3 GB
Q6_Kgguf1.4 GB2.3 GB3.1 GB4.2 GB6.3 GB
mlx-8bitmlx1.8 GB2.7 GB3.5 GB4.6 GB6.7 GB
Q8_0gguf1.8 GB2.7 GB3.5 GB4.6 GB6.7 GB

Run it

Copy a command for the runtime you use.

  • mlx-3bit
    • LM Studiolms get mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-3bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-3bit
  • mlx-4bit
    • LM Studiolms get mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-4bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-4bit
  • Q4_K_M
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF:Q4_K_M
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF@q4_k_m
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF:Q4_K_M
  • Q5_K_M
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF:Q5_K_M
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF@q5_k_m
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF:Q5_K_M
  • mlx-6bit
    • LM Studiolms get mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-6bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-6bit
  • Q6_K
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF:Q6_K
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF@q6_k
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF:Q6_K
  • mlx-8bit
    • LM Studiolms get mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-8bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Qwen-1.5B-8bit
  • Q8_0
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF:Q8_0
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF@q8_0
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF:Q8_0

Notes

Distilled from DeepSeek-R1 onto Qwen2.5-Math-1.5B (Qwen2.5 originally Apache 2.0); the card recommends temperature 0.6, no system prompt, and forcing responses to start with "<think>\n".

Questions

How much VRAM does DeepSeek R1 Distill Qwen 1.5B need?
At mlx-3bit, DeepSeek R1 Distill Qwen 1.5B needs about 2 GB of memory at 8k context and 6 GB at 128k, counting the weights, the KV cache and runtime buffers.
What is the best GPU for running DeepSeek R1 Distill Qwen 1.5B locally?
The cheapest hardware in the US that runs DeepSeek R1 Distill Qwen 1.5B fully in memory at 128k context is the NVIDIA Tesla P40, $345 at Amazon, at about 38 tok/s (estimated). The best value is the NVIDIA GeForce RTX 3080 10GB (about 83 tok/s for $469). The fastest we price is the NVIDIA GeForce RTX 5090, at about 197 tok/s.
Can I run DeepSeek R1 Distill Qwen 1.5B on a 16 GB GPU?
Yes: at Q8_0 it fits fully in 16 GB at 128k context. On the NVIDIA GeForce RTX 5080 it runs at about 105 tok/s (estimated).
Can I run DeepSeek R1 Distill Qwen 1.5B on a 24 GB GPU?
Yes: at Q8_0 it fits fully in 24 GB at 128k context. On the NVIDIA GeForce RTX 3090 Ti it runs at about 111 tok/s (estimated).
How do I run DeepSeek R1 Distill Qwen 1.5B locally?
With Ollama: ollama run hf.co/unsloth/DeepSeek-R1-Distill-Qwen-1.5B-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 1.5B released under?
MIT License, which allows commercial use.