# Llama 3.3 70B Instruct datasheet

> The classic 48 GB model. Two used 3090s or a large unified-memory box; bandwidth, not capacity, sets the speed.

Canonical: https://thelocalbench.com/models/llama-3.3-70b · Data as of 2026-10-12 · The Local Bench (thelocalbench.com), data as of 2026-10-12, CC BY 4.0

Flags: [verified] checked against primary source · [sourced] cited · [estimate] formula · [sample] placeholder

## Specifications

| Field | Value | Flag | Source |
|---|---|---|---|
| params_total | 70.55 B | verified | https://huggingface.co/api/models/unsloth/Llama-3.3-70B-Instruct |
| params_active | 70.55 B | verified | https://huggingface.co/api/models/unsloth/Llama-3.3-70B-Instruct |
| layers | 80 | verified | https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/blob/main/config.json |
| kv_heads | 8 | verified | https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/blob/main/config.json |
| q_heads | 64 | verified | https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/blob/main/config.json |
| head_dim | 128 | verified | https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/blob/main/config.json |
| hidden_size | 8192 | verified | https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/blob/main/config.json |
| context | 131072 tokens | verified | https://huggingface.co/unsloth/Llama-3.3-70B-Instruct/blob/main/config.json |
| license | Llama 3.3 Community License | verified | https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct/blob/main/LICENSE |

## Memory and quants (total = weights + F16 KV + overhead)

| Quant | bpw | File GB | At 8k | At 32k | At 128k | KLD | 24 GB card, 8k |
|---|---|---|---|---|---|---|---|
| Q3_K_M | 3.91 | 34.27 [verified] | 37.5 | 45.8 | 78.7 | 0.071 [sample] | offload |
| Q4_K_M | 4.85 | 42.52 [verified] | 45.8 | 54.0 | 87.0 | 0.017 [sample] | offload |
| Q5_K_M | 5.69 | 49.95 [verified] | 53.2 | 61.4 | 94.4 | 0.008 [sample] | offload |
| Q6_K | 6.56 | 57.89 [verified] | 61.1 | 69.4 | 102.3 | 0.004 [sample] | offload |
| Q8_0 | 8.5 | 74.98 [verified] | 78.2 | 86.5 | 119.4 | 0.001 [sample] | nofit |

## Speed (default quant, 8k context, single stream)

| Device | Verdict | Decode tok/s | Flag |
|---|---|---|---|
| [GeForce RTX 5060 Ti 16 GB](/hardware/rtx-5060-ti-16g.md) | nofit | — |  |
| [GeForce RTX 4060 Ti 16 GB](/hardware/rtx-4060-ti-16g.md) | nofit | — |  |
| [Radeon RX 7900 XTX 24 GB](/hardware/rx-7900-xtx-24g.md) | offload | 1.4 | estimate |
| [GeForce RTX 3090 24 GB (used)](/hardware/rtx-3090-24g-used.md) | offload | 1.4 | estimate |
| [CPU only · Ryzen 9 9950X + 128 GB DDR5](/hardware/cpu-ryzen-9-9950x-128g.md) | fit | 1.0 | estimate |
| [2× RTX 3090 (used) 48 GB](/hardware/2x-rtx-3090-48g-used.md) | tight | 12.4 | estimate |
| [Mac Studio M5 Max (base)](/hardware/mac-studio-m5-max.md) | nofit | — |  |
| [Ryzen AI Max+ 395 mini PC 128 GB](/hardware/ryzen-ai-max-395-128g.md) | fit | 2.5 | estimate |
| [GeForce RTX 4090 24 GB](/hardware/rtx-4090-24g.md) | offload | 1.4 | estimate |
| [NVIDIA DGX Spark FE 128 GB](/hardware/dgx-spark-128g.md) | fit | 2.7 | estimate |
| [MacBook Pro M4 Max 40c · 128 GB](/hardware/macbook-pro-m4-max-128g.md) | fit | 5.4 | estimate |
| [Mac Studio M5 Ultra (base)](/hardware/mac-studio-m5-ultra.md) | fit | 11.9 | estimate |
| [GeForce RTX 5090 32 GB](/hardware/rtx-5090-32g.md) | offload | 2.1 | estimate |
| [Mac Studio M5 Ultra · 512 GB](/hardware/mac-studio-m5-ultra-512g.md) | fit | 11.9 | estimate |

## Quality

LMArena text score: 1274.2 [sourced] (LMArena text, overall, entry 'llama-3.3-70b-instruct': 1274.2 (95% CI 1270.8–1277.7, 54,504 votes), leaderboard published 2026-10-08. CC-BY-4.0.)

JSON: https://thelocalbench.com/api/v1/models/llama-3.3-70b.json · Configurator: https://thelocalbench.com/configurator?model=llama-3.3-70b