Skip to content

Latest commit

 

History

History
77 lines (61 loc) · 3.22 KB

SPECS.md

File metadata and controls

77 lines (61 loc) · 3.22 KB

AI Model Requirements

Morgan uses two local models through Ollama: a light assistant model that stays loaded for conversation and simple PC actions, and a larger coding model that takes over coding jobs, web searches and multi-step tasks. Voice input is transcribed locally with Whisper, so neither model needs audio input.

Models

Role Model Size Native context Modalities
Assistant qwen3.5:4b ~3.3 GB 262,144 tokens Text and image
Coding qwen3.5:9b ~6.6 GB 262,144 tokens Text and image

Both models support tool calling. The assistant is kept loaded with the keep_alive setting. The coding model uses Ollama's default lifetime, so it is unloaded a few minutes after the last use.

There is no model selector: Morgan switches models by itself. A real coding job is handed to the coding model with the delegate_coding tool. When a request uses a web tool, becomes a supervised multi-step task, reaches four tool steps or two failed ones, or when the assistant calls the escalate tool, the rest of the request continues on the coding model with the same conversation, and the next request starts on the assistant again.

The base_model_name and coding_model entries of dev/core.json set the models, and the coding_model setting of the user configuration overrides the coding model; an empty coding_model falls back to the assistant model, which also turns escalation off.

Storage

  • Minimum free disk space: ~15 GB for both models and their model data
  • Recommended free disk space: 20+ GB to leave room for Ollama runtime data, updates and additional model data

Memory

The models do not define fixed minimum CPU, RAM or VRAM requirements for running through Ollama. Actual memory requirements depend on quantization, context length, KV cache configuration, multimodal inputs and hardware acceleration.

Practical requirements for Morgan

Minimum:

  • RAM: 16 GB
  • GPU VRAM: 8 GB
  • CPU: Modern 64-bit x86 CPU
  • GPU: Ollama-compatible GPU with hardware acceleration
  • Context: Reduced context may be required on lower-memory systems

Recommended:

  • RAM: 32 GB
  • GPU VRAM: 12–16 GB
  • CPU: Modern 8-core or better x86-64 CPU
  • GPU: Ollama-compatible GPU with 12+ GB VRAM
  • Storage: SSD/NVMe
  • Context: 32,768 tokens

Ideal:

  • RAM: 32–64 GB
  • GPU VRAM: 16 GB+
  • CPU: Modern high-performance 8-core or better x86-64 CPU
  • GPU: High-performance GPU supported by Ollama
  • Storage: NVMe SSD

The assistant and coding models occupy approximately 10 GB together, so a 16 GB GPU can keep both loaded with Morgan's recommended 32,768-token context while the coding model is in use. On a GPU with 8 to 12 GB of VRAM, Ollama swaps the two models, which adds a few seconds to the first reply after a coding job.

Both models accept images, so Morgan does not require a separate vision-language model. Morgan does not create derived Ollama models; when it starts Ollama it sets OLLAMA_CONTEXT_LENGTH to the context_length of the user configuration, which Settings edits (32,768 by default). An Ollama service that was already running keeps its own context length, and Morgan budgets against the context Ollama reports.