Inference
Dedicated capacity, not a shared queue
8 NVIDIA DGX Spark units in one cluster, serving open-weight models over an OpenAI-compatible API. You reserve throughput rather than discovering a rate limit in production.
Integration
Change the base URL
The API speaks the chat-completions shape your client already uses. Swapping provider is a configuration change — which is what makes the benchmark worth running before you commit to anything.
- Streaming and non-streaming
- Tool / function calling
- JSON and structured output
- Existing SDKs work unchanged
from openai import OpenAI
client = OpenAI(
base_url="https://api.aioctav.com/v1",
api_key=OCTAV_AI_KEY,
)
response = client.chat.completions.create(
model="glm-5.2",
messages=[{"role": "user", "content": "Summarise this filing."}],
)Models
What runs on the cluster
Open weights throughout. Bring a fine-tune and we host it as-is.
| Model | Family | Suited to |
|---|---|---|
| GLM-5.2 | Zhipu AI | General reasoning, long context, tool use |
| DeepSeek | DeepSeek AI | Code generation, maths, structured output |
| MiniMax-H3 | MiniMax | Long-form generation and multilingual work |
| Bring your own | Any open weights | We host and tune the model you already trust |
Why dedicated
What owning the hardware actually buys
OpenAI-compatible API
Point your existing client at a new base URL. If your code already talks to a chat-completions endpoint, it talks to ours.
Dedicated capacity
Reserve whole units rather than sharing a queue. Throughput you can plan around instead of a rate limit you discover in production.
Your data does not train anything
Prompts and completions are not retained for training and are not passed to a third-party provider. The hardware is ours.
Open weights, no lock-in
Every model we serve is open weights. If you leave, the model comes with you — the weights are not the product, the operation is.
Predictable cost
Reserved capacity is billed as capacity, not as a per-token meter that spikes the month a feature goes viral.
Run it in your own racks
The same stack can be installed on your hardware, in your building, with us operating it or handing it over.
Run your workload on it first
We benchmark with your prompts against whatever you use today, then quote against the result.