Updated minutes ago
Llama 3.2 3B
Meta · dense · 3.21B parameters · 131,072 context
Quality55.0
Architecture Details
TypeDENSE
Total Parameters3.21B
Active Parameters3.21B
Layers28
Hidden Dimension3,072
Attention Heads24
KV Heads8
Head Dimension128
Vocab Size128,256
Memory Requirements
BF16 Weights
6.4 GB
FP8 Weights
3.2 GB
INT4 Weights
1.6 GB
KV-Cache per Token114688 bytes
Activation Estimate0.50 GB
Fits on (single-node)
B200 SXM BF16B100 SXM BF16GB200 NVL72 (per GPU) BF16GB300 NVL72 (per GPU) BF16H200 SXM BF16H100 SXM BF16H100 PCIe BF16H100 NVL BF16
GPU Recommendations
RTX 4070 Tioptimal
BF16 · 1 GPU · vllm
100/100
score
Throughput
423.9 tok/s
Cost/Month
$237
Cost/M Tokens
$0.21
RTX 3080optimal
BF16 · 1 GPU · vllm
100/100
score
Throughput
639.2 tok/s
Cost/Month
$133
Cost/M Tokens
$0.08
RTX 4070 Superoptimal
BF16 · 1 GPU · vllm
100/100
score
Throughput
423.9 tok/s
Cost/Month
$209
Cost/M Tokens
$0.19
API Pricing Comparison
| Provider | Input $/M | Output $/M | Badges |
|---|---|---|---|
| together | $0.06 | $0.06 | Cheapest |
| fireworks | $0.10 | $0.10 |
Quality Benchmarks
MMLU63.4
HumanEval33.0
GSM8K68.0
MT-Bench73.0
Capabilities
Features
✓ Tool Use✗ Vision✓ Code✓ Math✗ Reasoning✓ Multilingual✓ Structured Output
Supported Frameworks
vllmsglangtgitensorrt-llmollama
Supported Precisions
BF16 (default)FP8INT4