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

# Boole AI Supported Models: Slugs, Pricing, and Performance

> Boole AI supports 42 compiler-optimized models for local and cloud use. Learn how to reference each model by slug in API calls and CLI commands.

Boole AI supports 42 models, each optimized for throughput by the nightly compiler. All models are available for both local binary deployment and the cloud API — you use the same model slug whether you are hitting `https://api.boole.dev/v1` or your local instance at `http://localhost:8000/v1`. Reference each model by its slug in API calls and CLI commands.

## Available Models

The table below lists the most commonly used models with their slugs, cloud API pricing, and median time-to-first-token (TTFT p50) measured on a single NVIDIA A10 GPU.

| Model                  | Slug                     | Input (per MTok) | Output (per MTok) | TTFT p50 |
| ---------------------- | ------------------------ | ---------------- | ----------------- | -------- |
| Llama 3.3 70B Instruct | `llama-3.3-70b-instruct` | \$0.09           | \$0.14            | 62 ms    |
| Qwen 2.5 72B Instruct  | `qwen-2.5-72b-instruct`  | \$0.11           | \$0.16            | 71 ms    |
| Mixtral 8x22B          | `mixtral-8x22b`          | \$0.13           | \$0.19            | 58 ms    |
| DeepSeek V3            | `deepseek-v3`            | \$0.08           | \$0.12            | 77 ms    |
| Whisper Large v3       | `whisper-large-v3`       | \$0.02/min       | —                 | 44 ms    |

<Note>
  Pricing is per million tokens for all text models, and per minute of audio for Whisper Large v3. The full list of all 42 available models is always current via `GET /v1/models`.
</Note>

## Referencing Models

Use the slug exactly as shown in the table above. The slug is identical across the cloud API, local API, and CLI.

**In an API call**, set the `model` field in your request body:

```json request.json theme={null}
{
  "model": "llama-3.3-70b-instruct",
  "messages": [{ "role": "user", "content": "Hello" }]
}
```

**In the local binary**, pass the slug to `--model`:

```bash theme={null}
./boole serve --model llama-3.3-70b-instruct
```

**To retrieve the live model list** at any time, call the models endpoint. This returns all 42 models with their current slugs, context lengths, and capabilities:

```bash theme={null}
curl https://api.boole.dev/v1/models \
  -H "Authorization: Bearer $BOOLE_API_KEY"
```

See the [Models API reference](/api-reference/models) for the full response schema.

## Model Optimization

Each model has its own independent optimization track. The nightly compiler runs a separate iteration loop per model, proposing and evaluating changes specific to that model's architecture — kernel shapes, quantization sensitivity, and batching behavior all vary by model family.

This means a compiler improvement for Llama 3.3 70B does not affect Qwen 2.5 72B, and vice versa. Each model's throughput and quality evolves on its own cadence based on what the compiler discovers.

Optimization logs for each individual model are published alongside the main crucible log at [github.com/boole-ai/crucible](https://github.com/boole-ai/crucible).

<Info>
  Prices shown are cloud API rates. Local deployment has zero per-token cost — once you have downloaded the binary and weights, you can run unlimited tokens with no additional charges.
</Info>