>>

Run Meta-Llama-3-70B on Lium

Meta · Text generation · 70.6B parameters · 141.1 GB BF16 weights · released 2024-04-17
It needs about 170 GB of VRAM. The cheapest setup that runs it is 4x RTX A6000 (192 GB), from $1.60 per hour, billed per second.

Other GPUs that run it

8x RTX 4080 SUPER 32GB256 GBfrom $2.40 / hourRent
B300 SXM6288 GBfrom $4.00 / hourRent

How to run it

Rent the pod above, open a terminal over SSH or Jupyter and serve meta-llama/Meta-Llama-3-70B:
vllm serve meta-llama/Meta-Llama-3-70B --tensor-parallel-size 4
Claude Code and Codex can do all of it through the Lium CLI.

Frequently asked questions

How much VRAM does Meta-Llama-3-70B need?

About 141.1 GB for the released BF16 weights, so plan on 170 GB with room for activations and KV cache at a short context. Longer contexts and bigger batches need more.

What GPU should I rent to run Meta-Llama-3-70B?

4x RTX A6000 (192 GB). On Lium that setup costs from $1.60 per hour, billed per second with no minimum.
All open models
For AI agents (CLI install, agent skill, headless signup): llms-full.txt