Models & pricing.
Everything below runs on the shared pool: call it with any API key, pay for the tokens you use, and never start or stop a GPU. The table is generated from the gateway's own catalog and rate card, so it is what the fleet actually serves and bills.
Shared models
Pass the model id in the request body — that is the string the gateway routes on, and the one your usage is recorded against.
| Model | Model id | Context | Input / 1M | Output / 1M |
|---|---|---|---|---|
| Qwen3 0.6B0.6B | qwen3-0.6b | 32K | $0.10 | $0.20 |
| Qwen3 4B Instructfine-tunable4B | qwen3-4b-instruct | 32K | $0.20 | $0.60 |
| Kimi K3MoE · 1.56 TB mxfp4 | kimi-k3 | 256K | $0.60 | $2.50 |
Prices are US dollars per million tokens. A request is rated on the usage block the gateway returns with it, summed over the billing period and rounded once at the invoice line — so a handful of small calls costs a fraction of a cent, not a rounded-up cent each.
What each one is for
qwen3-0.6bQwenQwen3 0.6B on the shared pool: a small, fast base model billed per token.
qwen3-4b-instructQwenQwen3 4B Instruct (2507) on the shared pool: the fine-tunable base — call it per token, or train it on your own dataset from the training tab.
kimi-k3KimiKimi K3 on the shared MI325X pool: 256K context with separate reasoning output and tool calling, billed per token. Accepts reasoning_effort of low, high or max — not medium.
Listing them from the API
The catalog is also available at runtime, which is the honest source if you are building a model picker — it includes any dedicated endpoints your project is running.
curl https://inference.legionedge.ai/v1/models \
-H "Authorization: Bearer $LEGIONEDGE_API_KEY"Dedicated GPU pricing
A dedicated endpoint rents whole GPUs instead of billing per token: you pay for the time the replica is up, whatever traffic it serves. Rates are per GPU, per hour.
| Hardware | Per GPU-hour |
|---|---|
MI325X-256GB | $4.79 |
A replica's hourly cost is this rate times the GPUs it needs — an eight-GPU model on MI325X bills eight times the row above for as long as it runs.