Llama 3.3 70B
The reference dense 70B. Its performance on every machine here is almost pure memory bandwidth divided by 40 GB, which makes it the cleanest possible test of a machine's decode capability. Superseded on quality, indispensable as a yardstick.
MetaDensegeneral16 machines can run it
Total parameters
70.554B
Official spec sets memory need
Active parameters
70.554B
Official spec sets decode speed
Context
128K
Official spec tokens
4-bit weights
38 GB
Calculated before KV cache
Model specification
Publisher
Meta Official spec
Architecture
Dense transformer Official spec
Total parameters
70.554B Official spec
Context length
131,072 tokens Official spec
Attention ⓘ
GQA · 8 KV heads × 128 dim × 80 layers Official spec
Licence
Llama 3.3 Community License Official spec source
Specification confidence ⓘ
High · verified 2026-09-07 Official spec source
Released
6 Dec 2024 Official spec
Official source
Quantizations and memory
| Quantization | Format | Bits/weight | Weights | Quality kept |
|---|---|---|---|---|
MLX 4-bit | mlx | 4.5 | 38.4 GB | 98.0% |
Q4_K_Mdefault | gguf | 4.85 | 41.0 GB | 98.5% |
Q8_0 | gguf | 8.5 | 71.2 GB | 99.9% |
BF16 | safetensors | 16 | 131.4 GB | 100.0% |
Weight sizes are computed from the parameter count and bits per weight plus a format-specific overhead for the layers that stay at higher precision — not read from a specific published file. Quality retention is an assumption, not a measured evaluation.
Recommendations
Cheapest that can run it
~0.5–0.7 t/s · €1.699
Lowest purchase price among configurations where the model fits at some quantization in our catalogue. Speed is not considered.
Cheapest above 20 t/s
~38.4–55.2 t/s · €6.499
20 tokens/second is roughly the point at which generation keeps pace with reading. Below it, interactive use feels like waiting.
Cheapest above 40 t/s
~38.4–55.2 t/s · €6.499
40 tokens/second is the threshold most people describe as comfortable for coding agents, where output arrives faster than you can review it.
Fastest with real measurements
Nothing in the database qualifies.
Highest throughput among configurations with an actual published measurement rather than our estimate.
Best throughput per euro
~38.4–55.2 t/s · €6.499
Highest decode tokens/second per EUR 1,000 of purchase price. Ignores running costs and resale — see the economics section for the full picture.
Hardware that runs Llama 3.3 70B0 measured, 16 estimated
16 of 16 rows
| Hardware↕ | Quant | Memory↕ | Decode▼ | Prefill↕ | Context | Price↕ | Fit | Confidence↕ |
|---|---|---|---|---|---|---|---|---|
| Dual RTX 5090 workstation (2x 32 GB) NVIDIA · 64 GB · 3,584 GB/s | MLX 4-bit | 43.5 GB | ~38.4–55.2 t/s | ~1420–2950 t/s | 16K | €6.499 | Comfortable | Estimated |
| Quad RTX 5090 workstation (4x 32 GB) NVIDIA · 128 GB · 7,168 GB/s | Q8_0 | 78.0 GB | ~32.3–46.4 t/s | ~1420–2950 t/s | 64K | €13.999 | Comfortable | Estimated |
| RTX PRO 6000 Blackwell workstation (96 GB) NVIDIA · 96 GB · 1,792 GB/s | Q4_K_M | 45.8 GB | ~24.9–35.8 t/s | ~461–958 t/s | 96K | €11.499 | Comfortable | Estimated |
| RTX PRO 6000 Max-Q workstation (96 GB, 300 W) NVIDIA · 96 GB · 1,792 GB/s | Q4_K_M | 45.8 GB | ~24.9–35.8 t/s | ~378–784 t/s | 96K | €11.999 | Comfortable | Estimated |
| Mac Studio M5 Ultra 96 GB Apple · 96 GB · 1,200 GB/s | Q4_K_M | 45.8 GB | ~19.7–28.3 t/s | ~329–683 t/s | 96K | €6.599 | Comfortable | Estimated |
| Mac Studio M5 Ultra 512 GB Apple · 512 GB · 1,200 GB/s | Q8_0 | 76.8 GB | ~11.3–16.3 t/s | ~411–854 t/s | 128K | €12.599 | Comfortable | Estimated |
| Mac Studio M5 Ultra 256 GB Apple · 256 GB · 1,200 GB/s | Q8_0 | 76.8 GB | ~11.3–16.3 t/s | ~411–854 t/s | 128K | €9.599 | Comfortable | Estimated |
| Mac Studio M5 Max 64 GB Apple · 64 GB · 614 GB/s | MLX 4-bit | 43.1 GB | ~11–15.8 t/s | ~206–427 t/s | 16K | €3.899 | Comfortable | Estimated |
| Mac Studio M3 Ultra 512 GB Apple · 512 GB · 819 GB/s | Q8_0 | 76.8 GB | ~6.3–9.1 t/s | ~99.2–206 t/s | 128K | €9.199 | Comfortable | Estimated |
| Mac Studio M3 Ultra 256 GB Apple · 256 GB · 819 GB/s | Q8_0 | 76.8 GB | ~6.3–9.1 t/s | ~99.2–206 t/s | 128K | €6.299 | Comfortable | Estimated |
| Mac Studio M5 Max 128 GB Apple · 128 GB · 614 GB/s | Q8_0 | 76.8 GB | ~5.8–8.4 t/s | ~206–427 t/s | 64K | €4.799 | Comfortable | Estimated |
| Mac mini M4 Pro 64 GB Apple · 64 GB · 273 GB/s | MLX 4-bit | 43.1 GB | ~3.9–5.7 t/s | ~17.2–35.8 t/s | 16K | €2.499 | Comfortable | Estimated |
| NVIDIA DGX Spark 128 GB NVIDIA · 128 GB · 273 GB/s | Q8_0 | 76.8 GB | ~2–2.9 t/s | ~177–368 t/s | 64K | €4.299 | Comfortable | Estimated |
| Framework Desktop (Ryzen AI Max+ 395, 128 GB) AMD · 128 GB · 256 GB/s | Q8_0 | 76.8 GB | ~1.3–1.9 t/s | ~48.9–102 t/s | 64K | €2.399 | Comfortable | Estimated |
| GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB) AMD · 128 GB · 256 GB/s | Q8_0 | 76.8 GB | ~1.2–1.7 t/s | ~48.9–102 t/s | 64K | €1.899 | Comfortable | Estimated |
| CPU-only workstation (Ryzen 9950X, 192 GB DDR5) Generic · 192 GB · 90 GB/s | Q8_0 | 76.8 GB | ~0.5–0.7 t/s | ~5.9–12.2 t/s | 128K | €1.699 | Comfortable | Estimated |
Similar models
Qwen3 32B
Alibaba Qwen · 32.762B · Apache 2.0
Gemma 4 31B
Google DeepMind · 31.273B · Gemma Terms of Use
Qwen3.8 27B
Alibaba Qwen · 27.781B · Apache 2.0
Gemma 3 27B
Google DeepMind · 27.432B · Gemma Terms of Use
Mistral Small 3.2 24B
Mistral AI · 24.011B · Apache 2.0
gpt-oss-120b
OpenAI · 116.829B (5.1B active) · Apache 2.0