Llama 3 · Llama-3.3 · HF: meta-llama/Llama-3.3-70B-Instruct
Llama 3.3 70B Instruct — VRAM, KV Cache & Hardware Matrix
Serverless API Rate
$0.71 in / $0.71 out per 1M
Parameters
70.6B Dense
Attention
GQA (8 KV / 64 Q)
Layers × Hidden
80 × 8192
Max Native Context
128k tokens
MoE Experts
N/A (Dense)
Quantization Memory Footprint
| Quantization | BPW | Weights Only | Total @ 8k Ctx | Total @ 32k Ctx | Quality |
|---|---|---|---|---|---|
| FP16 | 16.00 | 134.13 GiB | 138.51 GiB | 146.01 GiB | Lossless Reference |
| Q8_0 | 8.50 | 71.26 GiB | 75.63 GiB | 83.13 GiB | Near-Lossless |
| Q6_K | 6.57 | 55.08 GiB | 59.45 GiB | 66.95 GiB | High Precision |
| Q5_K_M | 5.69 | 47.70 GiB | 52.08 GiB | 59.58 GiB | Balanced High |
| Q4_K_M | 4.85 | 40.66 GiB | 45.03 GiB | 52.53 GiB | Recommended 4-bit |
| MLX_4BIT | 4.50 | 37.72 GiB | 42.10 GiB | 49.60 GiB | Apple Silicon Optimized |
| IQ4_XS | 4.25 | 35.63 GiB | 40.00 GiB | 47.50 GiB | Compact 4-bit |
| Q3_K_M | 3.91 | 32.78 GiB | 37.15 GiB | 44.65 GiB | Aggressive Compression |
Hardware Compatibility & Speed Leaderboard (Q4_K_M @ 32k Context)
| Hardware | Status | Total / Limit | Decode | TTFT (4k) | Max Safe Ctx | Inspect |
|---|---|---|---|---|---|---|
| Cloud NVIDIA H200 SXM (141 GB) | COMFORTABLE | 52.5 / 140.4 GiB | 74.1 tok/s | 0.88s | 128k | Simulate → |
| Cloud NVIDIA H100 SXM (80 GB) | COMFORTABLE | 52.5 / 79.4 GiB | 51.7 tok/s | 0.88s | 118k | Simulate → |
| Cloud NVIDIA A100 SXM (80 GB) | COMFORTABLE | 52.5 / 79.4 GiB | 30.7 tok/s | 2.87s | 118k | Simulate → |
| NVIDIA RTX PRO 6000 Blackwell (96 GB) | COMFORTABLE | 52.5 / 95.0 GiB | 27.7 tok/s | 3.57s | 128k | Simulate → |
| Apple Mac Studio M3 Ultra (256 GB) | COMFORTABLE | 52.3 / 199.7 GiB | 11.7 tok/s | 17.88s | 128k | Simulate → |
| Apple Mac Studio M3 Ultra (512 GB) | COMFORTABLE | 52.3 / 399.4 GiB | 11.7 tok/s | 17.88s | 128k | Simulate → |
| Apple Mac Studio M2 Ultra (192 GB) | COMFORTABLE | 52.3 / 149.8 GiB | 11.5 tok/s | 19.74s | 128k | Simulate → |
| Apple Mac Studio M4 Max (128 GB) | COMFORTABLE | 52.3 / 99.8 GiB | 7.7 tok/s | 27.23s | 128k | Simulate → |
| Apple MacBook Pro M3 Max (128 GB) | COMFORTABLE | 52.3 / 99.8 GiB | 5.6 tok/s | 36.48s | 128k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (128 GB) | COMFORTABLE | 52.3 / 96.0 GiB | 3.5 tok/s | 20.25s | 128k | Simulate → |
| Apple Mac Studio M4 Max (64 GB) | TUNABLE_UMA | 52.3 / 49.9 GiB | 7.7 tok/s | 27.23s | 45k | Simulate → |
| Apple MacBook Pro M3 Max (64 GB) | TUNABLE_UMA | 52.3 / 49.9 GiB | 5.6 tok/s | 36.48s | 45k | Simulate → |
| Apple Mac Mini M4 Pro (64 GB) | TUNABLE_UMA | 52.3 / 49.9 GiB | 3.7 tok/s | 56.29s | 45k | Simulate → |
| Ryzen AI Max+ 395 Strix Halo (64 GB) | TUNABLE_UMA | 52.3 / 48.0 GiB | 3.5 tok/s | 20.25s | 37k | Simulate → |
| Ryzen AI 9 HX 370 Mini-PC (64 GB) | TUNABLE_UMA | 52.3 / 48.0 GiB | 1.5 tok/s | 58.04s | 37k | Simulate → |
| NVIDIA RTX 5090 (32 GB) | OOM | 52.5 / 31.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 4090 (24 GB) | OOM | 52.5 / 23.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 5080 (16 GB) | OOM | 52.5 / 15.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 3090 (24 GB Used) | OOM | 52.5 / 23.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 5070 Ti (16 GB) | OOM | 52.5 / 15.0 GiB | OOM | — | OOM | Simulate → |
| AMD Radeon RX 7900 XTX (24 GB) | OOM | 52.5 / 23.0 GiB | OOM | — | OOM | Simulate → |
| Cloud NVIDIA L40S (48 GB) | OOM | 52.5 / 47.4 GiB | OOM | — | 16k | Simulate → |
| 2x NVIDIA RTX 3090 Rig (48 GB) | OOM | 52.5 / 46.6 GiB | OOM | — | 13k | Simulate → |
| NVIDIA RTX 5060 Ti (16 GB) | OOM | 52.5 / 15.0 GiB | OOM | — | OOM | Simulate → |
| NVIDIA RTX 3060 (12 GB) | OOM | 52.5 / 11.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 Pro (48 GB) | OOM | 52.3 / 37.4 GiB | OOM | — | 2k | Simulate → |
| Apple Mac Mini M4 (16 GB) | OOM | 52.3 / 10.0 GiB | OOM | — | OOM | Simulate → |
| Apple Mac Mini M4 (24 GB) | OOM | 52.3 / 18.0 GiB | OOM | — | OOM | Simulate → |
Interactive Llama 3.3 70B Instruct Simulator
70.6B Dense · GQA
1792 GB/s · 250 TFLOPS
3. Weight Quantization
4.85 bpw
4. KV Cache
32,768 tok
Compute Roofline
COMFORTABLE FIT Memory-Bandwidth Bound
Max Safe Ctx: 128k
Memory Allocation Footprint 52.53 / 95.00 GiB usable
Weights: 40.66 GiB
KV Cache: 10.00 GiB
Overhead: 1.88 GiB
Headroom: 42.47 GiB
Decode Speed
27.7
tok/s (batch=1)
Prefill Speed
1160
prompt tok/s
TTFT Latency
3.57s
for 4,096 tok
Active Params
70.6B
of 70.6B total
llama.cpp Server Command
llama-server -hf meta-llama/Llama-3.3-70B-Instruct:Q4_K_M -ngl 99 -fa on -c 32768 --port 8080| Context | KV Cache | Total VRAM | Decode | Status |
|---|---|---|---|---|
| 4k | 1.25 GiB | 43.22 GiB | 33.5 tok/s | COMFORTABLE |
| 8k | 2.50 GiB | 45.03 GiB | 32.5 tok/s | COMFORTABLE |
| 16k | 5.00 GiB | 47.53 GiB | 30.7 tok/s | COMFORTABLE |
| 32k | 10.00 GiB | 52.53 GiB | 27.7 tok/s | COMFORTABLE |
| 64k | 20.00 GiB | 62.53 GiB | 23.1 tok/s | COMFORTABLE |
| 128k | 40.00 GiB | 82.53 GiB | 17.4 tok/s | COMFORTABLE |
Buy-vs-Cloud-vs-API Monthly TCO Break-Even
USE SERVERLESS APIOwn Hardware/mo
$175.61
162.6h active + power
Cloud GPU/mo
$345.84
@$1.85/hr spot/pod
Serverless API/mo
$46.15
Payback: 230.2 mo
Recommended Hardware & Cloud Deployment
Contextual to fit statusEngineering FAQ — Llama 3.3 70B Instruct
How much VRAM does Llama 3.3 70B Instruct require in Q4_K_M?
Llama 3.3 70B Instruct weights occupy 40.66 GiB in Q4_K_M. Including 32k FP16 KV cache and runtime buffers, total memory is 52.53 GiB.
What is the cheapest local hardware that runs Llama 3.3 70B Instruct?
NVIDIA RTX PRO 6000 Blackwell (96 GB) (96 GB) runs Llama 3.3 70B Instruct Q4_K_M at ~27.7 tok/s decode and 3.57s TTFT.
How does Llama 3.3 70B Instruct scale KV cache at long context?
Llama 3.3 70B Instruct uses GQA with 8 KV heads across 80 layers (128 head_dim).