inference-v1
Systems Operational
High-Throughput Inference Runtime

Direct OpenAI-Compatible Endpoints with Zero Data Retention.

Low-latency, unmetered token generation powered by vLLM and PagedAttention kernels. Built for production agents and OpenRouter routing.

Base URL: https://api.voltinfra.xyz/v1
SLA: 99.9% Availability
Logging: Disabled (RAM Ephemeral)

Integration Snippet

Protocol: text/event-stream (SSE)
curl -X POST "https://api.voltinfra.xyz/v1/chat/completions" \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $VOLT_API_KEY" \
  -d '{
    "model": "Qwen/Qwen3-8B",
    "messages": [
      { "role": "system", "content": "You are a low-latency systems assistant." },
      { "role": "user", "content": "Return the cluster health spec in JSON." }
    ],
    "temperature": 0.2,
    "max_tokens": 512,
    "stream": true
  }'

Active Deployments

Runtime: vLLM (Marlin GEMM Kernel)
Model Target Qwen/Qwen2.5-8B-Instruct Production Active
Context Allocation 32,768 tokens (PagedAttention continuous batching)
Precision / Weights AWQ / INT4 quantized weights (Zero quality loss)
Output Throughput ~55 – 65 tokens/sec (stream generation)
Time to First Token (TTFT) ~400ms median
Health & Metrics GET /health → HTTP 200 OK

Transparent Pricing

Priced to parity with major cloud providers
Prompt (Input) Per 1M Tokens
$0.117 / 1M tokens

Raw context processing. Evaluated dynamically without minimum batching fee.

Completion (Output) Per 1M Tokens
$0.455 / 1M tokens

High-speed streaming generation with continuous token pipelining.

Zero Data Retention (ZDR) Policy

Volatile RAM Only
Context and activations exist strictly in ephemeral GPU memory buffers. All allocations are wiped upon socket closure.
Zero Disk Logs
Inference containers run with payload tracing permanently disabled. Prompts never touch persistent solid-state drives.
No Model Training
Your queries are strictly your own. Customer traffic is never pooled, inspected, or utilized for fine-tuning.