24 models live in us-east, us-west, eu-west3 regions
One API for 24 open-weight models. Pay per token on shared GPUs, then move to GPUs reserved for you when traffic grows, without changing a line of code.
For developers$5 in credit, no card
For teams at scaleDedicated GPUs, SLA, SSO
curl -N https://api.weir.example/v1/chat/completions \
-H "Authorization: Bearer $WEIR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "llama-3-3-70b",
"stream": true,
"messages": [
{ "role": "user", "content": "What does a KV cache do?" }
]
}'Responsellama-3-3-70b
Ready
A KV cache keeps the attention keys and values already computed for earlier tokens, so the model does not recompute them every time it generates a new one. Each decoding step then reads stored state instead of reprocessing the whole sequence, which keeps long responses fast and cheap.
Models
One key and one SDK for all of them, and switching between models on the same endpoint is a one-string change. Prices here are per million tokens, with no minimums and nothing billed while you are idle.
| Model | In / out | p50 |
|---|---|---|
| Llama 3.3 70B InstructMeta · chat | $0.54$0.72 | 212 ms |
| DeepSeek V3.1DeepSeek · chat · code | $0.55$1.65 | 280 ms |
| Qwen3 235B A22BAlibaba · chat · reasoning | $0.22$0.88 | 240 ms |
| DeepSeek R1 0528DeepSeek · reasoning | $0.70$2.40 | 520 ms |
| Qwen3 Coder 480B A35BAlibaba · code | $0.95$1.90 | 290 ms |
| Mistral Small 3.2 24BMistral AI · chat · vision | $0.10$0.25 | 110 ms |
Prices in USD per 1M tokens. p50 is the median time to first token over the last 24 hours.
Browse all 24 modelsArchitecture
Four hops between your code and the first token. On Llama 3.3 70B, the median trip takes 212 ms.
Your code sends one HTTPS request with a model name and messages, in the same shape for every model on the chat endpoint.
The closest of 3 regions ends TLS, checks the key and the rate limit, and turns bad requests away before they reach a GPU.
Sends the request to the region you pinned, or to the one with free capacity, batches it with others on the same model, and reuses cached prompt prefixes.
Shared H100s by default, or replicas reserved for you. The first token streams back and the rest follow as they decode.
Moving to dedicated GPUs changes one hop: the router sends your requests to replicas reserved for you. The request, the SDK and your code stay the same.
Two ways in
Both paths use the same API, the same SDKs and the same model names. What changes is whose GPUs answer and what the contract says.
Self-serve
A key, an SDK and a credit balance. Your first streamed reply takes about five minutes.
Enterprise
The same API on GPUs reserved for you, with the contract terms procurement asks for.
Pricing
Half input at $0.54, half output at $0.72, per 1M tokens. Billed per token, no minimums.
Dedicated H100s from $2.49 per GPU-hour, billed by the second.
See all pricingTrust
The answers procurement asks for, in one place, from the audit reports to the DPA.
Start on shared GPUs with $5 in credit. When you need capacity of your own, an engineer will size it with you.
For developers$5 in credit, no card
For teams at scaleDedicated GPUs, SLA, SSO