Grok-3
xAI · dense · 314B parameters · 131,072 context
Parameters
314B
Context Window
128K tokens
Architecture
Dense
Best GPU
H200 SXM
Cheapest API
$15.00/M
Quality Score
91/100
Intelligence Brief
Grok-3 is a 314B parameter DENSE model from xAI, featuring Grouped Query Attention (GQA) with 96 layers and 12,288 hidden dimensions. With a 131,072 token context window, it supports tools, vision, structured output, code, math, multilingual, reasoning. On standardized benchmarks, it achieves MMLU 92, HumanEval 78, GSM8K 95. The most cost-effective API deployment is via xai at $15.00/M output tokens. For self-hosted inference, H200 SXM delivers optimal throughput at $10211/month.
Architecture Details
Memory Requirements
BF16 Weights
628.0 GB
FP8 Weights
314.0 GB
INT4 Weights
157.0 GB
Fits on (single GPU) — most practical first
Fits on (multi-GPU with Tensor Parallelism)
Multi-GPU configurations use Tensor Parallelism (TP) to split model layers across GPUs. Requires NVLink or NVSwitch interconnect for optimal performance.
GPU Compatibility Matrix
Grok-3 is compatible with 7% of GPU configurations across 41 GPUs at 3 precision levels.
GPU Recommendations
FP8 · 4 GPUs · tensorrt-llm
80/100
score
Throughput
280.0 tok/s
Latency (ITL)
3.6ms
Est. TTFT
1ms
Cost/Month
$10211
Cost/M Tokens
$13.88
FP8 · 4 GPUs · tensorrt-llm
80/100
score
Throughput
280.0 tok/s
Latency (ITL)
3.6ms
Est. TTFT
1ms
Cost/Month
$3758
Cost/M Tokens
$5.11
FP8 · 2 GPUs · tensorrt-llm
78/100
score
Throughput
280.0 tok/s
Latency (ITL)
3.6ms
Est. TTFT
1ms
Cost/Month
$8541
Cost/M Tokens
$11.61
Deployment Options
API Deployment
xai
$15.00/M
output tokens
Single GPU
Requires multi-GPU setup (314 GB VRAM needed)
Multi-GPU
H200 SXM x4
280.0 tok/s
TP· $10211/mo
API Pricing Comparison
| Provider | Input $/M | Output $/M | Badges |
|---|---|---|---|
| xai | $3.00 | $15.00 | Cheapest |
Cost Analysis
| Provider | Input $/M | Output $/M | ~Monthly Cost |
|---|---|---|---|
| xaiBest Value | $3.00 | $15.00 | $90 |
Cost per 1,000 Requests
Short (500 tok)
$4.50
via xai
Medium (2K tok)
$18.00
via xai
Long (8K tok)
$54.00
via xai
Performance Estimates
Throughput by GPU
VRAM Breakdown (H200 SXM, FP8)
Precision Impact
bf16
157.0 GB
weights/GPU
fp8
78.5 GB
weights/GPU
~280.0 tok/s
Quality Benchmarks
Capabilities
Features
Supported Frameworks
Supported Precisions
Where to Deploy Grok-3
Self-Hosted Infrastructure
Similar Models
Grok-2
314B params · moe
Quality: 78
from $10.00/M
Grok 3
600B params · moe
Quality: 90
from $15.00/M
Nemotron 340B
340B params · dense
Quality: 85
from $4.20/M
DeepSeek Coder V2 236B
236B params · moe
Quality: 50
from $0.28/M
DeepSeek V2.5
236B params · moe
Quality: 78
from $0.28/M
Frequently Asked Questions
How much VRAM does Grok-3 need for inference?
Grok-3 requires approximately 628.0 GB of VRAM at BF16 precision, 314.0 GB at FP8, or 157.0 GB at INT4 quantization. Additional VRAM is needed for KV-cache (786432 bytes per token) and activations (~5.00 GB).
What is the best GPU for Grok-3?
The top recommended GPU for Grok-3 is the H200 SXM (x4) using FP8 precision. It achieves approximately 280.0 tokens/sec at an estimated cost of $10211/month ($13.88/M tokens). Score: 80/100.
How much does Grok-3 inference cost?
Grok-3 API inference starts from $3.00/M input tokens and $15.00/M output tokens. Self-hosted inference costs depend on your GPU configuration — use our ROI calculator for a detailed breakdown.