MAI-Code-1-Flash model frequently OOM-killed while Haiku/Sonnet succeed on same prompts

[baiwei0427]

Environment

• OS: Ubuntu 24.04, Azure VM (Standard_DS2_v2)
• RAM: 7.7 GB + 4 GB swap
• Available memory at time of test: ~5.5 GB
• Copilot CLI version: 1.0.60
• Node.js: v22.22.2

Problem

The mai-code-1-flash-internal model is frequently killed by the Linux OOM killer when running via Copilot CLI, while other models (claude-haiku-4.5, claude-sonnet-4.6) succeed on the exact same prompts under identical conditions.

Question

Is the higher memory consumption of mai-code-1-flash-internal compared to other models (e.g. claude-haiku-4.5) an expected/known behavior, or is this a bug? If expected, is there documentation on per-model memory requirements?

Steps to Reproduce

1. Clear page cache: sudo sh -c 'echo 3 > /proc/sys/vm/drop_caches'
2. Confirm ~5.5 GB available memory
3. Run: copilot -p 'Review myfile.py for bugs' --model mai-code-1-flash-internal --allow-all-tools
4. Process exits with SIGKILL (OOM) within 30-60 seconds
5. Run same prompt with --model claude-haiku-4.5 — succeeds

Observations

• Simple prompts (copilot -p 'say hello') work fine with MAI
• MAI fails consistently when using --allow-all-tools and reading files >200 lines
• Even without --allow-all-tools, piping file content (242 lines / 8KB) into the prompt causes OOM with MAI but not with Haiku
• After consecutive copilot invocations, residual copilot processes (~274 MB each) sometimes linger, compounding memory pressure
• Several background services reduce available memory but ~5.5 GB remains — which should be sufficient

Expected Behavior

MAI-Code-1-Flash should have similar or lower memory footprint compared to Claude Haiku, given that it is positioned as a lightweight/fast model.

Workaround

Using claude-haiku-4.5 or claude-sonnet-4.6 instead of MAI for code review tasks.