Llama 3.3 70B
A 71 billion-parameterHow many numbers the model learned during training — the usual rough measure of its size. dense model from Meta. A 70B instruct model with quality close to Llama 3.1 405B.
- Size
- 70.6Bdense
- Memory to run
- ~37 GBsmallest, 8K ctx
- Context
- 128Ktokens
- License
- RestrictedLlama 3.3 Community License
- Updated
- 18 Sept 2026
Which version to use
- Variant Llama 3.3 70B Instruct · MetaTuned to follow instructions and hold a conversation. The usual choice.
Where it runs
Best-fitting version on each reference system, at an 8K contextHow much text the model can consider at once, counted in tokens — roughly ¾ of a word each..
Runs well 7
- A100 80GB server~30 tok/s est.
- Dual RTX 309017.6 tok/s measured
- GB10 mini workstation~4 tok/s est.
- MacBook Pro M3 Max 64 GB7.4 tok/s measured
- MacBook Pro M4 Max 128 GB~8 tok/s est.
- Mac Studio M2 Ultra 192 GB12.1 tok/s measured
- Ryzen AI Max+ 395 mini PC~4 tok/s est.
Slowly (CPU or offload) 5
- CPU-only Ryzen 9 7950X~1 tok/s est.
- RTX 4060 Ti 16GB budget build~2 tok/s est.
- RTX 4090 workstation~4 tok/s est.
- RTX 5090 workstation~8 tok/s est.
- RX 7900 XTX desktop~4 tok/s est.
Too large 1
Benchmarks
Each benchmarkA fixed set of questions every model is given, so their scores can be compared on the same task. is developer-reported; prompts and settings differ between labs. Bars are relative to the best open result.
Variants & downloads
Each variant is a separate set of weights. Expand one for its downloads: quantizationStoring each of the model’s numbers with fewer bits, so the file is smaller and needs less memory. trades a little quality for a much smaller file, and the memory column adds the working memoryExtra space the model needs while it answers. It grows with the length of the conversation, on top of the file itself. a conversation needs on top.
Variant Llama 3.3 70B InstructRestricted
- What it is
- Tuned to follow instructions and hold a conversation. The usual choice.
- Publisher
- Meta
- License
- Llama 3.3 Community License · commercial use restricted
- Released
- 6 Dec 2024
- huggingface
- meta-llama/Llama-3.3-70B-Instruct
Contributions are not open yet, so there is nothing here from members.
| Quantization | Format | Bits | Download | Memory @ 8K | Publisher |
|---|---|---|---|---|---|
| Download BF16native | safetensors | 16 | 131 GiB | Metameta-llama/Llama-3.3-70B-Instruct | |
| Download Q4_K_Mk quant | gguf | 4.89 | 40.2 GiB | LM Studio Communitylmstudio-community/Llama-3.3-70B-Instruct-GGUF | |
| Download AWQ 4-bitawq | safetensors | 4.6 | 37.8 GiB | Unsloth AIunsloth/Llama-3.3-70B-Instruct-AWQ | |
| Download MLX 4-bitmlx | mlx | 4.5 | 37.0 GiB | MLX Communitymlx-community/Llama-3.3-70B-Instruct-4bit | |
| Download Q3_K_Mk quant | gguf | 3.91 | 32.1 GiB | Unsloth AIunsloth/Llama-3.3-70B-Instruct-GGUF |
Architecture details
- Parameters
- 70.55B
- Active / token
- All (dense)
- Architecture
- Dense
- Layers
- 80
- Attention heads
- 64
- KV heads
- 8
- Head dim
- 128
- KV cache @ 8K (fp16)
- 2.50 GiB
- Max context
- 131,072 tokens
Measured performance
Throughput on specific systems and runtimes, with the source of each measurement.
| Quantization | System | Runtime | Context | Measurements | Source |
|---|---|---|---|---|---|
| Llama 3.3 70B Instruct Q4_K_M (LM Studio Community) | Dual RTX 3090 (128 GB DDR5) | llama.cppcuda · b4600 | 4K | 390 tok/s · Prompt processing (512) 17.6 tok/s · Generation (128) | Mutinai illustrative fixtures |
| Llama 3.3 70B Instruct Q4_K_M (LM Studio Community) | MacBook Pro M3 Max 64 GB | llama.cppmetal · b4600 | 4K | 72 tok/s · Prompt processing (512) 7.4 tok/s · Generation (128) | Mutinai illustrative fixtures |
| Llama 3.3 70B Instruct MLX 4-bit (MLX Community) | MacBook Pro M4 Max 128 GB | MLX-LMmetal · 0.21.0 | 4K | 11.2 tok/s · Generation throughput 41 GB · Peak memory 110 tok/s · Prompt throughput 4,200 ms · Time to first token | Mutinai illustrative fixtures |
| Llama 3.3 70B Instruct Q4_K_M (LM Studio Community) | Mac Studio M2 Ultra 192 GB | llama.cppmetal · b4600 | 4K | 135 tok/s · Prompt processing (512) 12.1 tok/s · Generation (128) | Mutinai illustrative fixtures |
Community results
Reviews
Lineage
How this model’s variants relate to each other and to other models.
Sources & history
Sources
- Hugging Face Hub (live source, 1 record, 18 Sept 2026)
- Mutinai illustrative fixtures (illustrative fixture, 1 record, 13 Sept 2026)
External identifiers
- huggingface meta-llama/Llama-3.3-70B-Instruct
Field history
- license = "llama3.3" — Hugging Face Hub (current)
Timeline
- Llama 3.3 70B released