Health Test Model Attempts Failures | Generated by AI

Home 2026.06

Here’s what happened across all three attempts:

Health Test Summary

Attempt 1: unsloth/Qwen3-8B (4-bit, 7.5GB)

Attempt 2: Qwen/Qwen3.5-9B (cached, 18GB)

Attempt 3: Qwen/Qwen3-4B-FP8 (cached, 4.9GB)

What Actually Works (confirmed)

Root Problem

No compatible model is cached locally. The correct model (unsloth/Qwen3-4B-unsloth-bnb-4bit or unsloth/Qwen3-8B-unsloth-bnb-4bit) needs to be downloaded, but HF is slow right now.

Options

  1. Download overnight: huggingface-cli download unsloth/Qwen3-4B-unsloth-bnb-4bit — ~4GB, ~3-4hrs at current speed
  2. Use Qwen3-0.6B (tiny, fast download, good for pipeline validation)
  3. Use a model that’s already cached — but none of the cached ones (Qwen3.5-9B-VL, Qwen3-4B-FP8, Mistral-Small-3.2) are compatible with the Unsloth SFT pipeline as-is
  4. Wait for better network conditions

Want me to start the Qwen3-4B download in the background so it’s ready by tomorrow? Or try the 0.6B model for a quick pipeline validation first?


Back Donate