DeepSeek
DeepSeek LLM 67B (67B parameters) requires approximately 48.2 GB of VRAM with Q4_K_M quantization. For the best balance of quality and speed, we recommend hardware with at least 56 GB of VRAM.
Get started
— copy & paste to run locallyCopy-paste commands to run DeepSeek LLM 67B on your machine.
Run
docker run --rm -it ghcr.io/ggerganov/llama.cpp:full \
--hf-repo "deepseek-ai/deepseek-llm-67b-chat" \
--hf-file "deepseek-llm-67b-chat-Q4_K_M.gguf" \
-c 4096 -ngl 99Quick specs
About this model
Related models
Inference speed
Estimated decode speed (tokens/sec) for DeepSeek LLM 67B at Q4_K_M across popular GPUs and Apple Silicon, including multi-GPU rigs, using the fastest local runtime per device. Fastest is 2× RX 7900 XTX 24GB at ~20 tok/s. Speed is memory-bandwidth bound, so cards that fit the whole model in VRAM run far faster than ones that offload to system RAM.
| GPU / Mac | Memory | Quant | Speed (tok/s) | Fits? |
|---|---|---|---|---|
2× RX 7900 XTX 24GB | 48 GB | Q4_K_M | 19.5 | Heavy offload |
MacBook Pro M4 Max 128GB | 128 GB | Q4_K_M | 16.0 | Fits |
Mac Studio M3 Ultra 256GB | 256 GB | Q4_K_M | 14.8 | Fits |
MacBook Pro M4 Max 64GB | 64 GB | Q4_K_M | 12.4 | Heavy offload |
Mac Studio M2 Ultra 128GB | 128 GB | Q4_K_M | 12.3 | Fits |
Mac Studio M1 Ultra 128GB | 128 GB | Q4_K_M | 11.7 | Fits |
 2× RTX 4090 24GB | 48 GB | Q4_K_M | 10.3 | Heavy offload |
| 2× RTX 3090 24GB | 48 GB | Q4_K_M | 9.4 | Heavy offload |
| 4× RTX 3060 12GB | 48 GB | Q4_K_M | 8.3 | Heavy offload |
| RTX 5090 32GB | 32 GB | Q4_K_M | 6.2 | Too big |
MacBook Pro M4 Pro 48GB | 48 GB | Q4_K_M | 5.8 | Too big |
MacBook Pro M3 Max 64GB | 64 GB | Q4_K_M | 5.0 | Heavy offload |
MacBook Pro M1 Max 64GB | 64 GB | Q4_K_M | 4.5 | Heavy offload |
RX 7900 XTX 24GB | 24 GB | Q4_K_M | 2.9 | Too big |
| RTX 4090 24GB | 24 GB | Q4_K_M | 2.0 | Too big |
| RTX 4080 Super 16GB | 16 GB | Q4_K_M | 2.0 | Too big |
| RTX 3090 24GB | 24 GB | Q4_K_M | 2.0 | Too big |
| RTX 4070 12GB | 12 GB | Q4_K_M | 2.0 | Too big |
| RTX 3060 12GB | 12 GB | Q4_K_M | 2.0 | Too big |
| RTX 4060 8GB | 8 GB | Q4_K_M | 2.0 | Too big |
Estimates for single-stream decoding at Q4_K_M; real tokens/sec varies with prompt length, context, batch size, and runtime build. Prompt processing (prefill) is faster than the decode figures shown here.
Quick picks
Best hardware
Run this model
Quantization
How much VRAM DeepSeek LLM 67B (67B) needs at each GGUF quant, and whether it fits a 24 GB card (RTX 4090 / 3090). The recommended Q4_K_M uses ~40.9 GB — about 43% less VRAM than Q8_0, at a small quality cost.
| Quant | Bits | VRAM (weights) | Quality | Fits 24 GB? |
|---|---|---|---|---|
| Q2_K | 2 | 26.1 GB | Low | Too big |
| Q3_K_S | 3 | 32.8 GB | Low | Too big |
| NVFP4 | 4 | 37.5 GB | Medium | Too big |
| Q4_K_Mrecommended | 4 | 40.9 GB | Medium | Too big |
| Q5_K_M | 5 | 48.2 GB | High | Too big |
| Q6_K | 6 | 54.9 GB | High | Too big |
| Q8_0 | 8 | 71.7 GB | Very High | Too big |
| F16 | 16 | 137.4 GB | Maximum | Too big |
VRAM shown is quantized weights only; add ~1–3 GB runtime overhead plus KV cache for your context length. Lower quants trade quality for memory — Q4_K_M is the usual sweet spot; Q2/Q3 only when you must fit a bigger model.
Quality benchmarks
Coding
Reasoning
General
Source: official · 2024-01-05
Hardware compatibility
Computing compatibility...
Memory breakdown
Frequently asked questions
DeepSeek LLM 67B (67B parameters) requires approximately 48.2 GB of VRAM with Q4_K_M quantization. Lower quantizations like Q4_K_M use less memory but may reduce quality.
Yes, MacBook Pro M4 Max 96GB can run DeepSeek LLM 67B with a compatibility score of 58/100. It provides 96 GB of memory and achieves approximately 16.0 tokens per second.
The recommended quantization for DeepSeek LLM 67B is Q4_K_M, which offers the best balance between model quality and memory efficiency. Higher quantizations preserve more quality but require more VRAM.
The top recommended hardware for DeepSeek LLM 67B: NVIDIA H100 80GB (score: 66/100), NVIDIA H800 80GB (score: 65/100), NVIDIA A100 80GB (score: 64/100). These provide the best combination of memory, bandwidth, and compute for running this model locally.
Yes, DeepSeek LLM 67B is well-suited for chat as well as reasoning. It was designed with these use cases in mind.
See also