Llama 3.3 70B Instruct
Llama · Llama 3.3 Community License
Flagship Llama 3.3 model with best-in-class general and coding performance.
- Parameters
- 70.6B
- Architecture
- Dense
- Context
- 131,072 tokens
- Released
- 2024-12-06
- Engines
- llama.cpp, ollama, vLLM, TGI
- Builder Tools
- Cursor, Continue, Aider, Open WebUI, LM Studio
Parameters
70.6B
VRAM
61 GB
Context
128K
Formats
6
GPUs
22
Llama 3.3 70B Instruct (70.6B) requires 61 GB VRAM at recommended quality (Q6_K). On NVIDIA Grace Blackwell Ultra GB300, expect approximately 55 tok/s at Q8_0. For the best experience, Mac Studio AI Builder ($3,999) is recommended.
Source: OwnRig methodology
61 GB
Q6_K
52 GB
128K tokens
Chat
VRAM Requirements
| Quality | Quantization | VRAM | File Size |
|---|---|---|---|
| full | Q8_0 | 78 GB | 70 GB |
| recommended | Q6_K | 61 GB | 52 GB |
| recommended | Q5_K_M | 51 GB | 43 GB |
| efficient | Q4_K_M | 41 GB | 35 GB |
| compressed | Q3_K_M | 33 GB | 27 GB |
| compressed | Q2_K | 25.6 GB | 21 GB |
Context Length Impact
KV cache VRAM at Q6_K quality. Longer context = more memory.
| Context | KV Cache | Total VRAM |
|---|---|---|
| 2K | 1.2 GB | 62.2 GBexceeds 24 GB |
| 4K | 2.3 GB | 63.3 GBexceeds 24 GB |
| 8K | 4.6 GB | 65.6 GBexceeds 24 GB |
| 16K | 9.2 GB | 70.2 GBexceeds 24 GB |
| 32K | 18.4 GB | 79.4 GBexceeds 24 GB |
| 64K | 36.9 GB | 97.9 GBexceeds 24 GB |
| 128K | 73.7 GB | 134.7 GBexceeds 24 GB |
Compatible GPUs
22 devices| NVIDIA Grace Blackwell Ultra GB300 | Q8_0 | 55 tok/s | Excellent |
| Apple M4 Max (128GB Unified) | Q4_K_M | 18 tok/s | Acceptable |
| Apple M4 Max (64GB Unified) | Q3_K_M | 18 tok/s | Acceptable |
| Apple M4 Pro (48GB) | Q4_K_M | 12 tok/s | Acceptable |
| Apple M4 Ultra (192GB) | Q4_K_M | 24 tok/s | Acceptable |
| AMD Radeon Pro W7900 | Q4_K_M | 6 tok/s | Acceptable |
| NVIDIA RTX PRO 6000 Blackwell | Q8_0 | 14 tok/s | Acceptable |
| NVIDIA RTX PRO 6000 Blackwell Max-Q | Q8_0 | 13 tok/s | Acceptable |
| NVIDIA GeForce RTX 4090 | Q3_K_M | 6 tok/s | Marginal |
| NVIDIA GeForce RTX 5090 | Q4_K_M | 8 tok/s | Marginal |
| Apple M3 Pro (18GB Unified) | Q2_K | – | Not viable |
| NVIDIA GeForce RTX 3080 10GB | Q2_K | – | Not viable |
| NVIDIA GeForce RTX 4060 8GB | Q2_K | – | Not viable |
| NVIDIA RTX 4060 Laptop (40-60W) | Q2_K | – | Not viable |
| NVIDIA RTX 4070 Laptop (80-115W) | Q2_K | – | Not viable |
| NVIDIA GeForce RTX 4070 Ti 12GB | Q2_K | – | Not viable |
| AMD Radeon RX 7600 | Q3_K_M | – | Not viable |
| AMD Radeon RX 7900 XTX | Q3_K_M | – | Not viable |
| AMD Radeon RX 9070 | Q3_K_M | – | Not viable |
| AMD Radeon RX 9060 XT 16GB | Q3_K_M | – | Not viable |
| AMD Radeon RX 9060 XT 8GB | Q3_K_M | – | Not viable |
| NVIDIA GeForce RTX 5060 8GB | Q2_K | – | Not viable |
Showing 22 of 22 entries
Builder Context
Llama 3.3 70B Instruct is commonly used with Cursor, Continue, Aider, Open WebUI, LM Studio. For an AI coding workflow, pair it with an embedding model like nomic-embed-text for local RAG.
Recommended Builds
Complete PC builds that can run Llama 3.3 70B Instruct.
Frequently Asked Questions
- How much VRAM does Llama 3.3 70B Instruct need?
- Llama 3.3 70B Instruct requires 61 GB VRAM at recommended quality (Q6_K). At lower quality settings, it can fit in as little as 25.6 GB.
- What is the best GPU for Llama 3.3 70B Instruct?
- The NVIDIA Grace Blackwell Ultra GB300 delivers the best performance for Llama 3.3 70B Instruct, achieving 55 tok/s at Q8_0 with an excellent rating.
- What quantization should I use for Llama 3.3 70B Instruct?
- For the best quality, use Q6_K (61 GB VRAM). If your GPU has limited VRAM, Q2_K (25.6 GB) is the most efficient option with acceptable quality.
- Is Llama 3.3 70B Instruct good for coding?
- Yes. Llama 3.3 70B Instruct is used with Cursor, Continue, Aider, Open WebUI, LM Studio for local AI coding. For the best coding experience, pair it with an embedding model for local RAG.
Related Guides
Tutorial
The Complete Guide to Running LLMs Locally
Run large language models locally: hardware needs, Ollama and llama.cpp, model picks by use case, and quantization.
Explainer
VRAM: The Only Spec That Matters for AI
VRAM for local AI: what it is, why models need it, how quantization cuts requirements, and a VRAM table for major models.
Explainer
How we test: OwnRig's benchmark methodology
How OwnRig measures tokens per second, rates model compatibility, and keeps hardware data current. Our methodology, tools, and known limitations.
Data confidence: estimated. Source
VRAM requirements are calculated from model parameters and may vary by inference engine, context length, and batch size. Performance estimates are based on community benchmarks and should be verified for your specific configuration.Llama is a trademark of its respective owner. OwnRig is not affiliated with or endorsed by the model creator.