Inference for open models
Run GLM and Qwen on Astorias AI with your existing OpenAI code.
Every request lands on a stack you can inspect
Serve
Chat completions on a Hugging Face id or a checkpoint you bring.
Batch
JSONL in, a receipt out. Overnight jobs skip the interactive pool.
Fine-tune
LoRA on a hosted model, then serve the adapter or export it.
OpenAI-compatible
Same paths and payloads. Swap the base URL and keep your client.
Featured models
Hugging Face repo IDs on the chat endpoint. Open a row for the last 24 hours of tok/s.
| Model | Context | Quant | Input | Output | tok/s |
|---|---|---|---|---|---|
| 262128 | FP8 | $0.10 | $1.60 | 64 | |
| 1048576 | NVFP4 | $0.10 | $0.45 | 24 | |
100%
Uptime, last 7 days
Fine-tune any model
Any dataset. Serverless LoRA. You keep the weights.
- Pick a modelHugging Face id or a checkpoint you already host.
- Attach a datasetJSONL in object storage.
- Create the jobOne call. No cluster to rent.
- Pull the weightsExport safetensors, or serve the adapter on the same endpoint.
Before you integrate
Dedicated inference
Reserved capacity for your org. We optimize for your workload to give maximum savings.