> ## Documentation Index
> Fetch the complete documentation index at: https://docs.radium.cloud/llms.txt
> Use this file to discover all available pages before exploring further.

# Provenance

> What Radium models are built on and what we add

# Provenance

Radium is a managed inference platform, not a model lab. We serve fine-tuned and optimized variants of open-weight and licensed models, adding reliability, speed and safety on top.

## What models are behind Radium?

Radium offers three model families. Here's what we can disclose about their origins:

### hal-1.0

A general-purpose reasoning model optimized for coding, analysis and creative writing.

* **Base family:** State-of-the-art open-weight or licensed model
* **Fine-tuning:** Custom dataset for Radium optimization
* **Context window:** Up to 250,000 tokens
* **Key additions:**
  * Improved instruction following
  * Reduced verbose output

### clarke-1.0

A high-capability model with strong performance on reasoning, math and long-context tasks.

* **Base family:** State-of-the-art open-weight or licensed model
* **Fine-tuning:** Custom dataset for Radium optimization
* **Context window:** Up to 1,000,000 tokens
* **Key additions:**
  * Advanced reasoning pathways
  * Better document comprehension
  * Guardrail-tuned for safety without over-refusing

### tycho-1.0

A fast, cost-optimized model for high-throughput applications.

* **Base family:** State-of-the-art open-weight or licensed model
* **Fine-tuning:** Custom dataset for Radium optimization
* **Context window:** Up to 125,000 tokens
* **Key additions:**
  * Quantized (int8) serving for improved throughput and efficiency
  * Speculative decoding for reduced TTFT
  * Optimized for summarization, classification and extraction

## What Radium adds

Beyond the base model weights, Radium provides:

| Feature | Description |
| - | - |
| **Managed inference stack** | Owned GPU clusters, custom CUDA kernels, optimized scheduling |
| **Multi-region serving** | Canada (live) with United States planned |
| **Unified API** | One OpenAI-compatible endpoint for all models |
| **Automatic failover** | If one model is overloaded, traffic routes to the next best |
| **Usage analytics** | Per-key, per-model, per-endpoint cost and latency breakdown |
| **Enterprise controls** | SSO, RBAC and audit logs |
| **Custom fine-tuning** | Train on your data with our infrastructure |

## Why not disclose exact base models?

Radium models are continuously updated with the latest weights, optimizations and safety patches. Pinning them to a specific public model would create confusion as improvements roll out. We commit to:

* **Benchmark transparency:** All eval results are public and reproducible
* **Behavioral disclosure:** What the model can and cannot do
* **Version pinning:** Option to pin to a specific model version for reproducibility

## Model version history

| Version | Date | Key changes |
| - | - | - |
| hal-1.0-20260101 | Q2 2026 | Initial release |
| clarke-1.0-20260101 | Q2 2026 | Initial release |
| tycho-1.0-20260101 | Q2 2026 | Initial release |

Pin to a specific version:

```python theme={null}
response = client.chat.completions.create(
    model="hal-1.0",
    extra_headers={"X-Model-Version": "20260101"},
    messages=[{"role": "user", "content": "Hello"}],
)
```

Default behavior always routes to the latest stable version.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.