Llama 3.3 70B Instruct
Data checked: 2026-10-12. Review interval shown below. Prices need review after 30 days; specifications and benchmark collections after 90 days. Dates describe the recorded check, not a live retailer quote.
- Verified Checked against the primary source (for example the model’s config.json on Hugging Face).
- Sourced Taken from a cited source; not yet independently checked.
- Estimate Calculated with the published formulas; not a measurement.
- Sample Placeholder data for layout; not a real figure.
- Not yet measured No public, citable score found for this benchmark yet. We never fill the gap with a guess.
Verified metadata is not a measured performance result. Only cited runs matching model, device, quant and recorded context can replace speed estimates. Estimates assume one stream and F16 KV cache, with a ±25% uncertainty band. A published run applies only to its recorded runtime and conditions; missing settings are unknown.
Evidence labels · Freshness policy · Calculation methodology
At a glance
The classic 48 GB model. Two used 3090s or a large unified-memory box; bandwidth, not capacity, sets the speed.
At Q4_K_M and 8k context it needs about 45.8 GB: 42.5 GB of weights, 2.7 GB of KV cache and 0.6 GB of runtime overhead. It is dense, so every token reads all 42.5 GB of weights; speed scales with memory bandwidth.
Estimated setup
Quant | Q4_K_M |
|---|---|
Smallest standard card size at 8k | 96 GB |
Max context, 24 GB card | does not fit |
Runtimes | llama.cpp, Ollama, LM Studio, MLX |
Quantization options
Total memory = weights + KV cache (F16) + runtime overhead. File sizes marked Verified are the sum of GGUF files from bartowski/Llama-3.3-70B-Instruct-GGUF on the Hugging Face API. KV cache costs 0.328 MB per token before sliding-window savings.
Quant | bpw | File | At 8k | At 32k | At 128k | KL-divergence | 24 GB card, 8k |
|---|---|---|---|---|---|---|---|
| Q3_K_M | 3.91 | 34.3 GBVerified | 37.5 GB | 45.8 GB | 78.7 GB | 0.071Sample | Needs offload |
| Q4_K_M | 4.85 | 42.5 GBVerified | 45.8 GB | 54.0 GB | 87.0 GB | 0.017Sample | Needs offload |
| Q5_K_M | 5.69 | 50.0 GBVerified | 53.2 GB | 61.4 GB | 94.4 GB | 0.008Sample | Needs offload |
| Q6_K | 6.56 | 57.9 GBVerified | 61.1 GB | 69.4 GB | 102.3 GB | 0.004Sample | Needs offload |
| Q8_0 | 8.50 | 75.0 GBVerified | 78.2 GB | 86.5 GB | 119.4 GB | 0.001Sample | Does not fit |
Why this matters: What is quantization? · What is a KV cache?
Decode speed by device
Q4_K_M, single stream, 8k context. Estimates use the brief §7.3 formula (η × bandwidth ÷ bytes per token) and are ±25% until measured.
Speed table
Device | Verdict | Decode, tok/s | Range | Runtime |
|---|---|---|---|---|
| 2× RTX 3090 (used) 48 GB | Tight | 12.4Estimate | ±25% | llama.cpp (estimate) |
| Mac Studio M5 Ultra (base) | Fits | 11.9Estimate | ±25% | llama.cpp (estimate) |
| Mac Studio M5 Ultra · 512 GB | Fits | 11.9Estimate | ±25% | llama.cpp (estimate) |
| MacBook Pro M4 Max 40c · 128 GB | Fits | 5.4Estimate | ±25% | llama.cpp (estimate) |
| NVIDIA DGX Spark FE 128 GB | Fits | 2.7Estimate | ±25% | llama.cpp (estimate) |
| Ryzen AI Max+ 395 mini PC 128 GB | Fits | 2.5Estimate | ±25% | llama.cpp (estimate) |
| GeForce RTX 5090 32 GB | Needs offload | 2.1Estimate | ±25% | llama.cpp (estimate) |
| GeForce RTX 4090 24 GB | Needs offload | 1.4Estimate | ±25% | llama.cpp (estimate) |
| Radeon RX 7900 XTX 24 GB | Needs offload | 1.4Estimate | ±25% | llama.cpp (estimate) |
| GeForce RTX 3090 24 GB (used) | Needs offload | 1.4Estimate | ±25% | llama.cpp (estimate) |
| CPU only · Ryzen 9 9950X + 128 GB DDR5 | Fits | 1.0Estimate | ±25% | llama.cpp (estimate) |
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. Dataset
llama-perplexity --kl-divergence on a fixed corpus. Speed and quality are never merged into one score. Sample values are illustrative and must not be used to select a model.Architecture
Parameters, total | 70.55 BVerifiedsafetensors.total from the HF API. Read from an ungated mirror of the gated official repo; identical architecture. Source |
|---|---|
Parameters, active per token | 70.55 BVerifiedDense: all parameters active. Source |
Layers | 80VerifiedRead from an ungated mirror of the gated official repo; identical architecture. Source |
Query heads | 64VerifiedRead from an ungated mirror of the gated official repo; identical architecture. Source |
KV heads (GQA) | 8VerifiedRead from an ungated mirror of the gated official repo; identical architecture. Source |
Head dimension | 128VerifiedRead from an ungated mirror of the gated official repo; identical architecture. Source |
Hidden size | 8192VerifiedRead from an ungated mirror of the gated official repo; identical architecture. Source |
Operating conditions
Max context | 131,072 tokensVerifiedSource |
|---|---|
Licence | Llama 3.3 Community LicenseVerifiedSource |
Weights | meta-llama/Llama-3.3-70B-Instruct |
GGUF files | bartowski/Llama-3.3-70B-Instruct-GGUF |
Usable memory, discrete | VRAM − 0.5 GB2 |
Usable memory, unified | RAM × 0.752 |
Fits when | total ≤ 90% of usable2 |
1 Estimates: decode ≈ η × bandwidth ÷ (active weights + KV cache), η 0.60 discrete / 0.45 unified / 0.50 CPU (concept brief §7.3). Not measurements.
2 Fit rules from the concept brief §7.3, including the 0.75 unified-memory factor used by canirun.ai.
Flags: Verified = checked against config.json / HF API; Sourced = from a cited source; Estimate = formula; Sample = placeholder. Markdown · JSON · Report a problem