One command turns any open-source Hugging Face model into a production API. XAI1 selects the hardware, quantizes the weights, and tunes the CUDA kernels — you get an endpoint, billed per token.
No GPUs to provision. No CUDA to tune. No servers to babysit.
Any open-source model on Hugging Face — LLMs, Whisper-class ASR, TTS, vision, embeddings. Paste the model_id, or bring your own fine-tune.
The engine profiles the model, selects the cheapest hardware that meets your latency target, quantizes where it's lossless-in-practice, and compiles tuned CUDA kernels.
An OpenAI-compatible endpoint, live in minutes. Serverless autoscaling from zero to thousands of GPUs — billed per token, never for idle.
Same model, same silicon. The difference is what XAI1 does to it before your first request lands.
Per-token pricing breaks down outside LLMs, so XAI1 meters each modality in its own unit. All prices public. Batch −50%. Cached input −80%.
Auto-selected hardware never means surprise metering — the dashboard shows a pre-deploy estimate and a per-request cost breakdown. Here's the simple version.
Every price is public except enterprise contracts. Usage is metered per modality, and idle time is never billed — on any tier.
Run XAI1 in our cloud, your VPC, or your own metal. Compliance artifacts on request — no twelve-touch sales cycle required.
Talk to an engineer Response from a deploy engineer — not a sales deck — within one business day.Optimized snapshots keep warm-pool cold starts sub-second for most models under 20B parameters. Larger models use predictive pre-warming based on your traffic pattern; Pro's fast-lane class guarantees a warm floor.
Never, on any tier. Serverless endpoints scale to zero and bill per token (or audio-hour, character, image). Dedicated endpoints bill per GPU-second of active work only.
The engine profiles your model against a live fleet-price table and picks the cheapest configuration meeting your latency target. Before anything deploys, you see a cost estimate; after, every request carries a cost breakdown in the dashboard and API response headers.
Quantization is applied only when it passes automated quality evals against the full-precision baseline (perplexity + task-specific suites). You can pin full precision per endpoint with --precision fp16.
Prompts and outputs are never used for training, never shared, and retained only as long as your log window (24h Hobby, 30d Pro, custom Enterprise). VPC and self-host deployments keep data entirely inside your network.
Any model whose license permits your use case — the deploy flow surfaces the model's license (Apache-2.0, MIT, Llama Community, etc.) and flags restrictions before the endpoint goes live.