readygpu

DeepSeek V3.2

672B total, 38.5B active · 160k context · MIT License · released 2025-12-01

chatreasoningagentic

In short

DeepSeek V3.2 is an open model from DeepSeek with 672B parameters (38.5B active per token). At Q4_K_M, DeepSeek V3.2 needs about 379 GB of memory at 8k context and 383 GB at 64k, counting the weights, the KV cache and runtime buffers. Nothing we price in the US runs DeepSeek V3.2 fully in memory at 64k 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 US

Fully in memory at 64k 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
Mac Studio M3 Ultra 512GBQ4_K_Mfits12 tok/s$23,0000.5

39 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
Q4_K_Mgguf377.6 GB378.8 GB380.6 GB382.9 GB387.6 GB
mlx-5bitmlx430.3 GB431.6 GB433.4 GB435.7 GB440.4 GB
Q5_K_Mgguf443.5 GB444.7 GB446.5 GB448.8 GB453.5 GB
Q6_Kgguf513.5 GB514.7 GB516.5 GB518.8 GB523.5 GB
Q8_0gguf664.3 GB665.6 GB667.3 GB669.7 GB674.3 GB

Run it

Copy a command for the runtime you use.

  • Q4_K_M
    • Ollamaollama run hf.co/unsloth/DeepSeek-V3.2-GGUF:Q4_K_M
    • LM Studiolms get unsloth/DeepSeek-V3.2-GGUF@q4_k_m
    • llama.cppllama-server -hf unsloth/DeepSeek-V3.2-GGUF:Q4_K_M
  • mlx-5bit
    • LM Studiolms get mlx-community/DeepSeek-V3.2-mlx-5bit
    • MLXmlx_lm.server --model mlx-community/DeepSeek-V3.2-mlx-5bit
  • Q5_K_M
    • Ollamaollama run hf.co/unsloth/DeepSeek-V3.2-GGUF:Q5_K_M
    • LM Studiolms get unsloth/DeepSeek-V3.2-GGUF@q5_k_m
    • llama.cppllama-server -hf unsloth/DeepSeek-V3.2-GGUF:Q5_K_M
  • Q6_K
    • Ollamaollama run hf.co/unsloth/DeepSeek-V3.2-GGUF:Q6_K
    • LM Studiolms get unsloth/DeepSeek-V3.2-GGUF@q6_k
    • llama.cppllama-server -hf unsloth/DeepSeek-V3.2-GGUF:Q6_K
  • Q8_0
    • Ollamaollama run hf.co/unsloth/DeepSeek-V3.2-GGUF:Q8_0
    • LM Studiolms get unsloth/DeepSeek-V3.2-GGUF@q8_0
    • llama.cppllama-server -hf unsloth/DeepSeek-V3.2-GGUF:Q8_0

Notes

No Jinja chat template is included; use the provided Python encoding scripts (encoding_dsv32.py). Recommended sampling: temperature 1.0, top_p 0.95. Sparse or compressed attention: KV read per token is estimated as full attention. Excludes 13.4B parameters of multi-token-prediction layers, which local runtimes do not load.

Questions

How much VRAM does DeepSeek V3.2 need?
At Q4_K_M, DeepSeek V3.2 needs about 379 GB of memory at 8k context and 383 GB at 64k, counting the weights, the KV cache and runtime buffers.
Can I run DeepSeek V3.2 on a 16 GB GPU?
No: it does not fit in 16 GB at 64k context, even with system RAM.
Can I run DeepSeek V3.2 on a 24 GB GPU?
No: it does not fit in 24 GB at 64k context, even with system RAM.
How do I run DeepSeek V3.2 locally?
With Ollama: ollama run hf.co/unsloth/DeepSeek-V3.2-GGUF:Q4_K_M (the Q4_K_M build). LM Studio and llama.cpp commands are on this page.
What licence is DeepSeek V3.2 released under?
MIT License, which allows commercial use.