Twai AI Inference & Cloud Savings Dashboard

Hybrid Cloud + Local Architecture

Gemini Cloud Orchestration + Qwen 3.8 Local Inference

High-level planning, user interaction, and tool governance are managed by Google Gemini in the cloud, while deep code synthesis, continuous refactoring, and heavy context processing are dispatched locally to nvidia/Qwen3.8-27B-NVFP4 on your dedicated NVIDIA RTX 4500 rig.

Engine vLLM 0.31

Precision NVFP4 / FP8 KV

Context 32,768 Tok

                        Total Cloud Cost Avoided
                        
$3.34 USD
≈ R$ 17.86 BRL

                        Cost Avoided
                        vs Gemini 1.5/2.5 Pro pricing

                        Total Tokens Offloaded
                        
890,059
858,155 Prompt / 31,904 Gen

                        Heavy Context
                        ≈ 1,780 document pages

                        Inference Operations
                        
71 calls
Local Tasks Executed

                        Orchestrated
                        Avg Quota Rem: ~97.1%

                        Prefix Cache Hit Rate
                        
81.3%
697.8k cached / 160.4k prefilled

                        ~35 tokens/s
                        Sub-second response initiation

Cloud Quotas Remaining (Available Models)

Real-time daily capacity preserved by offloading heavy reasoning to Qwen 3.8

Daily Pool • Resets at 00:00 UTC

Google Gemini

Gemini 3.8 Flash (Orchestrator)

Orchestrator

95.3%

Remaining

1,429 / 1,500 daily requests Active

OpenAI GPT

GPT-4o / o3-mini

Offloaded

97.2%

Remaining

486 / 500 daily requests Standby

Anthropic Claude

Claude 3.5 / 3.7 Sonnet

Offloaded

98.9%

Remaining

989 / 1,000 daily requests Standby

Cloud Cost Comparison

What you would have paid across top commercial AI providers

Token Distribution

Prefill context efficiency and generation breakdown

vLLM Telemetry

Financial Savings Matrix

Official token pricing vs your local execution cost on Twai

Provider / Model	Input / 1M	Output / 1M	Cloud Cost (USD)	Cloud Cost (BRL)	Local Rig Cost	Net Savings

Local Rig Infrastructure

Host Name: Twai (Local LAN / Tailscale)
Hardware Accelerator: NVIDIA RTX 4500 (24GB VRAM)
Tailscale Address: 100.111.95.99
Electricity Consumption: ~0.03 kWh (< $0.01 USD)

vLLM Engine Configuration

Model Architecture: nvidia/Qwen3.8-27B-NVFP4
Quantization Scheme: NVFP4 Weights + FP8 KV Cache
Context Window: 32,768 Tokens
Prefix Caching: Enabled (81.3% Hit Rate)