readygpu

DeepSeek R1 Distill Llama 70B

70.6B parameters · 128k context · Llama 3.3 License · released 2025-01-20

chatreasoningcoding

In short

DeepSeek R1 Distill Llama 70B is an open model from DeepSeek with 70.6B parameters. At mlx-3bit, DeepSeek R1 Distill Llama 70B needs about 32 GB of memory at 8k context and 32 GB at 8k, counting the weights, the KV cache and runtime buffers. Nothing we price in the UK runs DeepSeek R1 Distill Llama 70B fully in memory at 8k context at a usable speed; see the table for what comes closest.

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

What to run it on in the UK

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
GMKtec EVO-X2 Ryzen AI Max+ 395 128GBQ8_0fits2 tok/s£3,3200.5
Mac Studio M4 Max 64GBQ4_K_Mfits8 tok/s£3,3992.3
Mac Studio M3 Ultra 96GBQ5_K_Mfits10 tok/s£5,4991.8
Mac Studio M3 Ultra 256GBQ8_0fits7 tok/s£8,4990.8
Mac Studio M3 Ultra 512GBQ8_0fits7 tok/s£17,4950.4
NVIDIA RTX PRO 6000 BlackwellQ8_0fits14 tok/s£29,5040.5
Mac Studio M4 Max 128GBQ8_0fits5 tok/s––
Mac mini M4 Pro 64GBQ4_K_Mfits4 tok/s––
Framework Desktop Ryzen AI Max+ 395 128GBQ8_0fits2 tok/s––
NVIDIA DGX Spark 128GBQ8_0fits2 tok/s––
ASUS Ascent GX10 128GBQ8_0fits2 tok/s––
Framework Desktop Ryzen AI Max+ 395 64GBQ4_K_Mfits3 tok/s––
NVIDIA Tesla P40Q4_K_Moffload2 tok/s£2617.7
Intel Arc Pro B60Q4_K_Moffload2 tok/s£7422.7
AMD Radeon RX 7900 XTQ4_K_Moffload2 tok/s£7892.4
AMD Radeon RX 7900 XTXQ4_K_Moffload2 tok/s£9992.3
AMD Radeon AI PRO R9700Q4_K_Moffload3 tok/s£1,6741.9
NVIDIA GeForce RTX 3090Q4_K_Moffload2 tok/s£1,7071.3
NVIDIA GeForce RTX 3090 TiQ4_K_Moffload2 tok/s£2,1991.0
NVIDIA GeForce RTX 5090Q4_K_Moffload4 tok/s£5,0000.8
NVIDIA GeForce RTX 4090Q4_K_Moffload2 tok/s£6,0150.4

19 other cards and machines 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-3bitmlx28.7 GB32.0 GB39.6 GB49.8 GB70.2 GB
mlx-4bitmlx37.0 GB40.2 GB47.9 GB58.0 GB78.4 GB
Q4_K_Mgguf39.6 GB42.8 GB50.5 GB60.7 GB81.1 GB
Q5_K_Mgguf46.5 GB49.8 GB57.4 GB67.6 GB88.0 GB
mlx-6bitmlx53.4 GB56.6 GB64.3 GB74.5 GB94.8 GB
mlx-8bitmlx69.8 GB73.1 GB80.7 GB90.9 GB111.3 GB
Q8_0gguf69.8 GB73.1 GB80.7 GB90.9 GB111.3 GB

Run it

Copy a command for the runtime you use.

  • mlx-3bit
    • LM Studiolms get mlx-community/DeepSeek-R1-Distill-Llama-70B-3bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Llama-70B-3bit
  • mlx-4bit
    • LM Studiolms get mlx-community/DeepSeek-R1-Distill-Llama-70B-4bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Llama-70B-4bit
  • Q4_K_M
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:Q4_K_M
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF@q4_k_m
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:Q4_K_M
  • Q5_K_M
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:Q5_K_M
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF@q5_k_m
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:Q5_K_M
  • mlx-6bit
    • LM Studiolms get mlx-community/DeepSeek-R1-Distill-Llama-70B-6bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Llama-70B-6bit
  • mlx-8bit
    • LM Studiolms get mlx-community/DeepSeek-R1-Distill-Llama-70B-8bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-R1-Distill-Llama-70B-8bit
  • Q8_0
    • Ollamaollama run hf.co/unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:Q8_0
    • LM Studiolms get unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF@q8_0
    • llama.cppllama-server -hf unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:Q8_0

Notes

Derived from Llama-3.3-70B-Instruct, so the Llama 3.3 license applies, although the card header says MIT. Avoid system prompts, use temperature 0.5-0.7, and force responses to start with "<think>\n".

Questions

How much VRAM does DeepSeek R1 Distill Llama 70B need?
At mlx-3bit, DeepSeek R1 Distill Llama 70B needs about 32 GB of memory at 8k context and 32 GB at 8k, counting the weights, the KV cache and runtime buffers.
Can I run DeepSeek R1 Distill Llama 70B on a 16 GB GPU?
No: it does not fit in 16 GB at 8k context, even with system RAM.
Can I run DeepSeek R1 Distill Llama 70B on a 24 GB GPU?
Only partly: at 8k context it does not fit in 24 GB, so part of it runs from system RAM and it slows to about 2 tok/s on the NVIDIA Tesla P40 (estimated).
How do I run DeepSeek R1 Distill Llama 70B locally?
With Ollama: ollama run hf.co/unsloth/DeepSeek-R1-Distill-Llama-70B-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 Llama 70B released under?
Llama 3.3 License.