Co-Residence Guide¶

Running multiple models simultaneously on a single Mac specializes workloads — use a fast model for simple queries and a powerful model for complex reasoning. This page shows which model pairs fit in 64 GB unified memory and how to configure them.

Memory budget¶

On a 64 GB Mac, the usable memory for models is approximately 55–58 GB after accounting for macOS overhead (~5–6 GB) and other running applications.

Model memory requirements¶

Pairing guide¶

Pairs That Fit Comfortably (≤ 55 GB total)¶

Pairs That Fit Tightly (55–58 GB total)¶

Pairs That Do NOT Fit¶

Recommended pairings¶

The "Developer" Stack¶

# Two models running simultaneously
models:
  - name: gemma4:e4b-nvfp4
    role: Fast chat / quick queries
    port: 11434
    context: 8192  # Short context for speed

  - name: qwen3-coder:30b
    role: Code generation / complex tasks
    port: 11435
    context: 32768  # Longer context for code

Total memory: ~42 GB — leaves 13 GB for macOS and headroom.

The "Reasoning" Stack¶

models:
  - name: gemma4:e2b-nvfp4
    role: Fast chat / simple queries
    port: 11434
    context: 8192

  - name: gemma4:26b-nvfp4
    role: Deep reasoning / analysis
    port: 11435
    context: 32768

Total memory: ~26 GB — leaves 29 GB for macOS and headroom.

The "MoE Power" Stack¶

models:
  - name: gemma4:e4b-nvfp4
    role: Fast chat
    port: 11434
    context: 8192

  - name: qwen3.6:35b-a3b-coding-nvfp4
    role: MoE coding / complex tasks
    port: 11435
    context: 32768

Total memory: ~32 GB — leaves 23 GB for macOS and headroom.

Context length tradeoffs¶

Reducing context length frees KV cache memory, allowing tighter pairs to fit:

Running multiple models¶

With Ollama¶

Ollama keeps loaded models in memory by default. To run two models simultaneously:

# Terminal 1: Start first model
ollama run gemma4:e4b-nvfp4

# Terminal 2: Start second model (in another terminal)
ollama run qwen3-coder:30b

Ollama will keep both in memory as long as they fit. If memory runs low, it will unload the least recently used model.

With MLX Engines¶

Each MLX engine instance loads its model independently:

# Terminal 1: Serve first model on port 8080
omlx serve --model gemma4:e4b-nvfp4 --port 8080

# Terminal 2: Serve second model on port 8081
omlx serve --model qwen3-coder:30b --port 8081

With Open WebUI¶

Open WebUI can manage multiple models and route requests based on your selection:

1. Configure both models in Open WebUI's model list
2. Select the appropriate model for each conversation
3. Open WebUI keeps both loaded and routes requests to the active model

Monitoring memory usage¶

# Check total memory pressure
memory_pressure

# Check per-process memory (Ollama)
ps aux | grep ollama

# Check per-process memory (MLX)
ps aux | grep mlx

# macOS activity monitor (GUI)
open -a "Activity Monitor"

What was learned¶

1. Gemma 4 e4b-nvfp4 pairs with everything — its 9.3 GB budget leaves room for any other model
2. DeepSeek R1 14B is the pairing challenge — its 40 GB KV cache dominates memory; reduce context to 32K to make it pair-friendly
3. The best all-around pair is gemma4:e4b-nvfp4 + qwen3-coder:30b — fast chat + excellent coding, ~42 GB total
4. You can run 3+ small models — gemma4:e2b-nvfp4 (6.8 GB) + gemma4:e4b-nvfp4 (9.3 GB) + gemma4:12b-nvfp4 (9.2 GB) = ~25 GB total
5. Always account for KV cache — model size alone is misleading; DeepSeek R1's 9 GB model needs 49 GB total at 128K context