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

# Browse and select Darkbloom inference models

> Use GET /v1/models to see which models are online, compare capabilities and pricing, and choose the right model for your workload and budget.

`GET /v1/models` returns the list of models currently available on the Darkbloom network. The response follows the OpenAI models list format, with additional fields for trust level, provider count, and hardware metadata. All listed models have been verified by the coordinator and have at least one active provider online.

## Request

```bash curl theme={null}
curl https://api.darkbloom.dev/v1/models \
  -H "Authorization: Bearer eigeninference-your-key-here"
```

## Response format

<ResponseField name="object" type="string">
  Always `"list"`.
</ResponseField>

<ResponseField name="data" type="object[]">
  Array of model objects.

  <Expandable title="model object fields">
    <ResponseField name="id" type="string">
      The model ID to pass in inference requests.
    </ResponseField>

    <ResponseField name="object" type="string">
      Always `"model"`.
    </ResponseField>

    <ResponseField name="owned_by" type="string">
      The organization that released the model weights.
    </ResponseField>

    <ResponseField name="provider_count" type="number">
      Number of active providers currently serving this model.
    </ResponseField>

    <ResponseField name="trust_level" type="string">
      Highest trust level among active providers: `self_signed` or `hardware`.
    </ResponseField>

    <ResponseField name="attested" type="boolean">
      `true` if at least one provider serving this model has passed recent attestation.
    </ResponseField>

    <ResponseField name="display_name" type="string">
      Human-readable model name for display purposes.
    </ResponseField>
  </Expandable>
</ResponseField>

## Available models

The catalog currently includes the following models. All are quantized to 8-bit for efficient Apple Silicon inference.

### Qwen3.5 27B Claude Opus 8-bit

**Model ID:** `qwen3.5-27b-claude-opus-8bit`

A 27-billion-parameter dense model distilled from Claude Opus. Delivers frontier-quality reasoning at a fraction of the compute cost of the full Opus model. Well-suited for complex reasoning, analysis, and code generation tasks that benefit from extended thinking.

| Property         | Value              |
| ---------------- | ------------------ |
| Architecture     | 27B dense          |
| Quantization     | 8-bit              |
| Min provider RAM | 36 GB              |
| Input price      | \$0.10 / 1M tokens |
| Output price     | \$0.78 / 1M tokens |

### Gemma 4 26B 8-bit

**Model ID:** `mlx-community/gemma-4-26b-a4b-it-8bit`

Google's Gemma 4 in a 26-billion-parameter mixture-of-experts configuration with only 4 billion parameters active per forward pass. Fast and memory-efficient, with multimodal instruction following. A good default for general-purpose tasks where cost and latency matter.

| Property         | Value               |
| ---------------- | ------------------- |
| Architecture     | 26B MoE, 4B active  |
| Quantization     | 8-bit               |
| Min provider RAM | 36 GB               |
| Input price      | \$0.065 / 1M tokens |
| Output price     | \$0.20 / 1M tokens  |

### Trinity Mini 8-bit

**Model ID:** `mlx-community/Trinity-Mini-8bit`

A 27-billion-parameter adaptive mixture-of-experts model optimized for agentic use cases — tool use, multi-step reasoning, and long-context tasks. The adaptive routing keeps active parameter count low while maintaining quality on structured tasks.

| Property         | Value            |
| ---------------- | ---------------- |
| Architecture     | 27B Adaptive MoE |
| Quantization     | 8-bit            |
| Min provider RAM | 48 GB            |

### Qwen3.5 122B MoE 8-bit

**Model ID:** `mlx-community/Qwen3.5-122B-A10B-8bit`

The highest-quality model in the catalog. 122 billion total parameters with 10 billion active per token — delivering near-full-model quality at significantly reduced inference cost. Best for tasks where output quality is the primary constraint.

| Property         | Value                |
| ---------------- | -------------------- |
| Architecture     | 122B MoE, 10B active |
| Quantization     | 8-bit                |
| Min provider RAM | 128 GB               |
| Input price      | \$0.13 / 1M tokens   |
| Output price     | \$1.04 / 1M tokens   |

### MiniMax M2.5 8-bit

**Model ID:** `mlx-community/MiniMax-M2.5-8bit`

A state-of-the-art coding and reasoning model with 239 billion total parameters and 11 billion active per token. Achieves approximately 100 tokens per second on Apple Silicon, making it competitive with much smaller models on throughput while delivering top-tier coding quality.

| Property         | Value                |
| ---------------- | -------------------- |
| Architecture     | 239B MoE, 11B active |
| Quantization     | 8-bit                |
| Min provider RAM | 256 GB               |
| Input price      | \$0.06 / 1M tokens   |
| Output price     | \$0.50 / 1M tokens   |

## Choosing a model

<AccordionGroup>
  <Accordion title="General-purpose tasks">
    Start with `mlx-community/gemma-4-26b-a4b-it-8bit`. It has the lowest output cost, runs on the widest range of provider hardware (36 GB+), and is fast enough for interactive use.
  </Accordion>

  <Accordion title="Complex reasoning and analysis">
    Use `qwen3.5-27b-claude-opus-8bit` for tasks requiring multi-step logic, careful analysis, or nuanced writing. The Claude Opus distillation gives it reasoning depth beyond its parameter count.
  </Accordion>

  <Accordion title="Best possible output quality">
    Use `mlx-community/Qwen3.5-122B-A10B-8bit`. With 122B total parameters, it produces the highest quality output in the catalog across most benchmarks.
  </Accordion>

  <Accordion title="Coding and software tasks">
    Use `mlx-community/MiniMax-M2.5-8bit`. It was trained for coding tasks and achieves approximately 100 tokens per second, which makes it practical for long code generation.
  </Accordion>

  <Accordion title="Agentic and tool-use workflows">
    Use `mlx-community/Trinity-Mini-8bit`. Its adaptive MoE routing is tuned for the structured reasoning patterns that appear in tool-use and multi-step agent loops.
  </Accordion>
</AccordionGroup>

<Note>
  The `provider_count` field tells you how many providers are currently online for each model. A count of zero means the model is in the catalog but no providers are serving it right now — your request will queue. The coordinator retries up to three times before returning an error.
</Note>