// Llama · Text LLM
Llama-3.1-70B-Instruct: VRAM & local deployment
70.6B parameters. About 50.2 GB at 4-bit and 158 GB at BF16 with one 8K-token request.
released 2024-07-16licence: Commercial with conditionsmemory figures are estimates
01Specs
| Publisher | Meta · meta-llama |
|---|---|
| Family / series | Llama · Llama-3.1 |
| Type | Text LLM (text-generation) |
| Parameters | 70.6B Source: counted from official safetensors weights |
| Context length | 128K tokens (131,072)max_position_embeddings in config.json |
| Architecture | 80 layers · GQA |
| Released weights | BF16 · 141 GBdownload |
| Licence | Llama 3.1 Community · Commercial with conditionsCommercial use allowed under the publisher’s own licence terms (user caps, acceptable-use policy, attribution or region limits). Read it before shipping. |
| Released | 2024-07-16 (Hugging Face repo created) |
| Access | Gated: accept the licence on Hugging Face before downloading. |
02How much VRAM it needs (estimate)
Total = weights + 10% + 0.5GB overhead + KV cache for one request (FP16 KV). More concurrent users multiply the KV column.
| Precision | Weights | Total @ 8K | Total @ 32K | Total @ 128K | Smallest GPU | Mac |
|---|---|---|---|---|---|---|
| BF16 / FP16 (original) | 141 GB | 158 GB | 166 GB | 199 GB | 4 × RTX PRO 6000 (384GB) | Mac 256GB |
| FP8 / INT8 / Q8_0 | 75.0 GB | 85.6 GB | 93.7 GB | 126 GB | RTX PRO 6000 96GB | Mac 128GB |
| 4-bit (Q4_K_M / AWQ / GPTQ-Int4) | 42.8 GB | 50.2 GB | 58.3 GB | 90.5 GB | 2 × RTX 5090 (64GB) | Mac 128GB |
KV cache per request: 8K 2.7 GB · 32K 10.7 GB · 128K 43.0 GB
Bits per weight: BF16 16, FP8 8.5, 4-bit 4.85. “Fits” = total ≤ 90% of memory; Mac GPU can use ~2/3 of unified memory (≤36GB) or ~3/4 (larger). For MoE, all experts must be in memory, so size by total parameters. Method →
03How to run it
Only runtimes the publisher or the runtime’s official library names for this model.
- Transformers — mentioned in the official model card
- Ollama — in the official Ollama library · llama3.1:70b
ollama run llama3.1:70b04Official repositories & quantised builds
2 official repo(s) from meta-llama.
- Llama-3.1-70B-Instruct Instruct
- Llama-3.1-70B main
05Other Llama models
All 16 →| Model | Params | 4-bit total | Released |
|---|---|---|---|
| Llama-3.1-8B-Instruct · same series | 8.03B | 6.9 GB | 2024-07-18 |
| Llama-3.1-405B · same series | 405.9B | 275 GB | 2024-07-16 |
| Llama-Prompt-Guard-2-22M | 71M | 0.5 GB | 2025-04-28 |
| Llama-Prompt-Guard-2-86M | 279M | 0.7 GB | 2025-04-28 |
| Llama-Guard-4-12B | 12B | 8.5 GB | 2025-04-23 |
| Llama-4-Scout-17B-16E-Instruct | 108.6B | 73.0 GB | 2025-04-02 |
| Llama-4-Maverick-17B-128E-Instruct | 401.6B | 268 GB | 2025-04-01 |
| Llama-3.3-70B-Instruct | 70.6B | 50.2 GB | 2024-11-26 |
| Llama-Guard-3-1B | 1.5B | 1.5 GB | 2024-09-20 |
| Llama-Guard-3-11B-Vision | 10.7B | 7.6 GB | 2024-09-20 |
06Sources
Data checked 2024-12-15. Parameters, context and licence are copied from the official repository; memory figures are our estimates.