Llama 3.3 70B
70.6B parameters · 128k context · llama3.3 · released 2024-11-26
chatcoding
What to run it on in the US
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.
| Hardware | Quant | Fits | Decode | Price | tok/s per 1,000 |
|---|---|---|---|---|---|
| Mac Studio M4 Max 128GB | Q8_0 | fits | 5 tok/s | – | – |
| Mac mini M4 Pro 64GB | Q4_K_M | fits | 4 tok/s | – | – |
| Framework Desktop Ryzen AI Max+ 395 128GB | Q8_0 | fits | 2 tok/s | – | – |
| Mac Studio M3 Ultra 512GB | Q8_0 | fits | 7 tok/s | – | – |
| NVIDIA DGX Spark 128GB | Q8_0 | fits | 2 tok/s | – | – |
| GMKtec EVO-X2 Ryzen AI Max+ 395 128GB | Q8_0 | fits | 2 tok/s | – | – |
| Mac Studio M3 Ultra 256GB | Q8_0 | fits | 7 tok/s | – | – |
| Mac Studio M4 Max 64GB | Q4_K_M | fits | 8 tok/s | – | – |
| Mac Studio M3 Ultra 96GB | Q5_K_M | fits | 10 tok/s | – | – |
| NVIDIA RTX PRO 6000 Blackwell | Q8_0 | fits | 14 tok/s | – | – |
| ASUS Ascent GX10 128GB | Q8_0 | fits | 2 tok/s | – | – |
| Framework Desktop Ryzen AI Max+ 395 64GB | Q4_K_M | fits | 3 tok/s | – | – |
| NVIDIA GeForce RTX 3090 | Q4_K_M | offload | 2 tok/s | – | – |
| NVIDIA GeForce RTX 5090 | Q4_K_M | offload | 4 tok/s | – | – |
| NVIDIA GeForce RTX 4090 | Q4_K_M | offload | 2 tok/s | – | – |
| AMD Radeon RX 7900 XTX | Q4_K_M | offload | 2 tok/s | – | – |
| AMD Radeon AI PRO R9700 | Q4_K_M | offload | 3 tok/s | – | – |
| AMD Radeon RX 7900 XT | Q4_K_M | offload | 2 tok/s | – | – |
| NVIDIA GeForce RTX 3090 Ti | Q4_K_M | offload | 2 tok/s | – | – |
| Intel Arc Pro B60 | Q4_K_M | offload | 2 tok/s | – | – |
| NVIDIA Tesla P40 | Q4_K_M | offload | 2 tok/s | – | – |
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).
| Quant | File | @8k | @32k | @64k | @128k |
|---|---|---|---|---|---|
| Q4_K_Mgguf | 39.6 GB | 42.8 GB | 50.5 GB | 60.7 GB | 81.1 GB |
| mlx-4bitmlx | 37.0 GB | 40.2 GB | 47.9 GB | 58.0 GB | 78.4 GB |
| Q5_K_Mgguf | 46.5 GB | 49.8 GB | 57.4 GB | 67.6 GB | 88.0 GB |
| mlx-8bitmlx | 69.8 GB | 73.1 GB | 80.7 GB | 90.9 GB | 111.3 GB |
| Q8_0gguf | 69.8 GB | 73.1 GB | 80.7 GB | 90.9 GB | 111.3 GB |
Run it
Copy a command for the runtime you use.
- Q4_K_M
- Ollama
ollama run hf.co/unsloth/Llama-3.3-70B-Instruct-GGUF:Q4_K_M - LM Studio
lms get unsloth/Llama-3.3-70B-Instruct-GGUF@q4_k_m - llama.cpp
llama-server -hf unsloth/Llama-3.3-70B-Instruct-GGUF:Q4_K_M
- Ollama
- mlx-4bit
- LM Studio
lms get mlx-community/Llama-3.3-70B-Instruct-4bit - MLX
mlx_lm.server --model mlx-community/Llama-3.3-70B-Instruct-4bit
- LM Studio
- Q5_K_M
- Ollama
ollama run hf.co/unsloth/Llama-3.3-70B-Instruct-GGUF:Q5_K_M - LM Studio
lms get unsloth/Llama-3.3-70B-Instruct-GGUF@q5_k_m - llama.cpp
llama-server -hf unsloth/Llama-3.3-70B-Instruct-GGUF:Q5_K_M
- Ollama
- mlx-8bit
- LM Studio
lms get mlx-community/Llama-3.3-70B-Instruct-8bit - MLX
mlx_lm.server --model mlx-community/Llama-3.3-70B-Instruct-8bit
- LM Studio
- Q8_0
- Ollama
ollama run hf.co/unsloth/Llama-3.3-70B-Instruct-GGUF:Q8_0 - LM Studio
lms get unsloth/Llama-3.3-70B-Instruct-GGUF@q8_0 - llama.cpp
llama-server -hf unsloth/Llama-3.3-70B-Instruct-GGUF:Q8_0
- Ollama