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

# Hardware

> GPU VRAM requirements and model sizing.

Bumblebee uses Ollama for local inference. VRAM requirements depend on which models you run and how large a context window you need.

## GPU VRAM guide

<Tabs>
  <Tab title="Minimum (~8 GB)">
    **Example GPUs:** RX 7600 8 GB, RTX 3050 8 GB, Arc A770 8 GB

    Smaller or quantized models only. Use aggressive quantization or point reflex at `gemma4:e4b` in entity YAML to lighten load. CPU-only via Ollama works for experiments but expect slow turns.
  </Tab>

  <Tab title="Recommended (~16 GB)">
    **Example GPUs:** RTX 4060 Ti 16 GB, RTX 4070 Ti Super 16 GB, RX 6800 XT 16 GB

    The common target for the default Bumblebee stack. Runs `gemma4:26b` for both reflex and deliberate reasoning (same weights, one model loaded). The soma noise engine reuses the reflex model at zero extra VRAM cost. Close other GPU-heavy apps if near the limit.
  </Tab>

  <Tab title="Comfortable (24+ GB)">
    **Example GPUs:** RTX 3090 24 GB, RTX 4090 24 GB, RX 7900 XTX 24 GB

    Comfortable dual-model setups with headroom. Room for larger context windows, higher thinking budgets, or a dedicated noise model without juggling VRAM.
  </Tab>

  <Tab title="Spacious (32+ GB)">
    **Example GPUs:** RTX 5090 32 GB, professional cards

    Room for larger deliberate models, extended context (64K+), or separate deliberate weights alongside the reflex model.
  </Tab>
</Tabs>

## Default models

| Model              | Role                         | Approx. VRAM |
| ------------------ | ---------------------------- | ------------ |
| `gemma4:26b`       | Reflex + deliberate chat     | \~16 GB      |
| `nomic-embed-text` | Memory similarity embeddings | \~274 MB     |

The embedding model loads on demand alongside the chat model — there is no separate embedding service. Both reflex and deliberate use the same weights with different token budgets, so only one model needs to be loaded at a time.

## Optional models

| Model        | Role                  | When to use                                                                      |
| ------------ | --------------------- | -------------------------------------------------------------------------------- |
| `gemma4:e4b` | Fast reflex layer     | Tight VRAM (\~8 GB setups). Set as `cognition.reflex_model` in entity YAML.      |
| `gemma3:1b`  | Dedicated noise model | Different character of inner voice. Set in `soma.noise.model`. Costs extra VRAM. |

## Context window and VRAM

Larger context windows use more memory. The default `max_context_tokens: 32768` (32K) is a good balance for 16 GB cards.

```yaml theme={null}
cognition:
  max_context_tokens: 65536   # 64K — needs more VRAM headroom
```

## Ollama settings

<Tip>For single-GPU setups, these defaults prevent overcommitting. The `ollama:reset` npm script sets them automatically.</Tip>

```env theme={null}
OLLAMA_MAX_LOADED_MODELS=1
OLLAMA_KEEP_ALIVE=60s
OLLAMA_CONTEXT_LENGTH=16384
OLLAMA_NUM_PARALLEL=1
```

```bash theme={null}
npm run ollama:reset
```

## MoE note

Gemma 4 uses a Mixture-of-Experts architecture. Active parameters per token are lower than the full model size, so actual VRAM usage during inference can be less than the raw parameter count suggests. Real-world fit depends on context length, thinking budget, quantization level, and concurrent platform activity.