Skip to main content

Frequently Asked Questions (FAQ)

Can I run GoingMerry completely offline?​

Yes. GoingMerry is built to operate under air-gapped security boundaries. Once weights are downloaded to your local registry cache, the execution engine runs offline. No telemetry, prompts, or vectors are transmitted outside your device.

How do I change the default listener port?​

By default, the server binds to 127.0.0.1:11434. You can bind to a custom host and port by defining the MERRY_HOST environment variable before launching the server:

export MERRY_HOST=127.0.0.1:8080
merry serve

What models does GoingMerry support?​

GoingMerry supports a wide variety of state-of-the-art model families. This includes:

  • Llama 4
  • Gemma 4
  • DeepSeek-V3
  • Phi 4
  • Qwen 3.0
  • Mistral Large
  • Hermes Agent

You can also import any GGUF or Safetensors model weights directly using a custom Modelfile.

Does GoingMerry support multi-GPU setups?​

Yes. GoingMerry automatically detects multiple graphics cards (e.g. SLI/NVLink Nvidia GPUs or unified memory clusters) and shards model layers across them to maximize throughput. If you wish to isolate execution to a specific GPU, configure environment flags:

# Force execution only on GPU 0
export CUDA_VISIBLE_DEVICES=0
merry serve

Why is GoingMerry faster than traditional runtimes?​

Standard runtimes often rely on generic builds designed for maximum CPU/GPU model compatibility, which introduces instruction branching and memory pre-fetching inefficiencies. GoingMerry compiles low-level C++ tensors with targeted AVX-512 vector instruction paths, organization structures that mirror CPU L3 cache lines, and direct memory pipelines that reduce cold-start latency to ~70ms.

Where are model weights stored on my disk?​

  • macOS: ~/.merry/models
  • Linux: /usr/share/merry/.merry/models or ~/.merry/models
  • Windows: C:\Users\<username>\.merry\models

You can redirect this storage target by changing the MERRY_MODELS environment variable.