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

# Models

> Supported inference models and their requirements

# Supported Models

Blindference currently supports three inference models across cloud and local backends.

## Model Reference

| Model ID                       | Provider | Backend   | VRAM / Requirements         | Tier | Speed    | Quality     |
| ------------------------------ | -------- | --------- | --------------------------- | ---- | -------- | ----------- |
| `groq:llama-3.3-70b-versatile` | Groq     | Cloud API | `GROQ_API_KEY` env var      | 0    | Fast     | High        |
| `gemini:gemini-2.5-flash`      | Google   | Cloud API | `GOOGLE_API_KEY` env var    | 0    | Fast     | Medium-High |
| `facebook/opt-125m`            | Local    | vLLM      | 0.5GB+ VRAM, `vllm` package | 0    | Variable | Dev/Testing |

## Resolution Order

When a job arrives, the node queries backends in registration order:

1. **vLLM** (if GPU available and model supported)
2. **Groq** (if `GROQ_API_KEY` set)
3. **Gemini** (if `GOOGLE_API_KEY` set)
4. **Mock** (always available, universal fallback)

The first backend that (a) is available and (b) advertises the requested `model_id` wins.

## Cloud API Setup

### Groq

```bash theme={null}
export GROQ_API_KEY="gsk_..."
```

Get your key at [console.groq.com](https://console.groq.com).

### Google Gemini

```bash theme={null}
export GOOGLE_API_KEY="AI..."
```

Get your key at [ai.google.dev](https://ai.google.dev).

## Local GPU Setup

### Minimum Requirements

* NVIDIA GPU with 0.5GB+ VRAM
* CUDA 11.8+ or ROCm (AMD)
* Python 3.10+

### Install vLLM

```bash theme={null}
pip install vllm
```

### Verify

```bash theme={null}
blindference-node models test --backend vllm --model facebook/opt-125m --prompt "Hello"
```

## Determinism

| Backend | Determinism Method                               |
| ------- | ------------------------------------------------ |
| vLLM    | `temperature=0`, `seed=42`, `enforce_eager=True` |
| Groq    | `temperature=0`, `seed=42` (API-native)          |
| Gemini  | `temperature=0` (best-effort)                    |
| Mock    | SHA-256 of `(model_id, prompt)`                  |

All cloud backends use a lightweight `[seed_anchor:{hash}]` prefix to reduce variance without restricting response creativity.

## Adding Custom Models

See [Model Backends](/compute/backends) for the full pluggable backend system.