Connecting our pool to OpenRouter.
We are currently working to connect our GPU pool to OpenRouter as an inference provider. Integration is in progress; a live provider listing has not been announced.
Retrivium’s role
Our role is to operate the GPU infrastructure and serve language models as an inference provider. OpenRouter supplies the model access and routing layer for applications.
Current stage
The current work is the integration of our compute pool in the provider role. Integration in progress is distinct from a production launch; this site does not announce an active public provider listing, a launch date or production traffic.
How the roles fit together
Retrivium focuses on the compute and model-serving layer. OpenRouter connects applications to model providers through its API and routing system. The integration follows this provider model.
Planned launch offer
Preparing for launch. These are proposed specifications and prices, subject to validation on production machines. The profiles are not enabled and their mapping to the OpenRouter model must be confirmed during onboarding.
- Model on OpenRouter
- qwen/qwen3.5-9b
- Source model
- Qwen/Qwen3.5-9B
- Total context, input + output
- 262,144
- Maximum output, including reasoning
- 81,920
- Input / 1M tokens
- $0.080
- Output / 1M tokens
- $0.130
- Target capabilities
- Text, streaming, tools, JSON / JSON Schema, reasoning
At the full output limit, up to 180,224 input tokens. A shorter requested output leaves more room for input within the total context.
Separate precision profiles
qwen/qwen3.5-9b-fp8FP8qwen/qwen3.5-9b-4bit4-bit — exact format pending
Both profiles have the same proposed prices. There is no announced cache discount or separate reasoning surcharge. BF16 is a future option, not part of this launch offer.
Planned locations
Production inference uses only machines owned and operated by Retrivium. EU endpoints guarantee routing, input processing and inference exclusively within the European Union, including load balancing, retries and failover. If EU capacity is unavailable, requests are rejected instead of being sent outside the EU. Actual serving countries are published for each endpoint.
Data policy
The launch policy prohibits training, persistent storage of inference content and individual request journals. Only aggregate usage statistics may be retained. Temporary RAM/GPU processing and memory-only caching are described in the policy. These requirements must be verified before activation.
Read the inference data policy →
Operator: IE Aleksandr Golubev, Tbilisi, Georgia. golubev.cto@gmail.com