Fizgig v5 Full Fine-Tuning: Train MiniMax H3 and Krea 2 on 16 GB

Fizgig v5.0.0 released on August 29 adds full fine-tuning of the MiniMax H3 (33B) and Krea 2 (12.9B) base models on a single consumer GPU, down to 16 GB VRAM (GitHub | release notes, Apache-2.0). Until now every Fizgig run trained a LoRA adapter on a frozen model. This release trains the model itself: full-rank weight updates with no adapter and no rank bottleneck, using a rotating trainable window over a 4-bit NF4 frozen base, with the bf16 master copy held in system RAM.

Fizgig: a train, fine-tune, repair and explore workbench for MiniMax H3, Krea 2 and Flux 2 Klein 9B

What your card can fine-tune

Every confirmed number comes from measured runs, not estimates: 8.8 to 12.3 GB peaks on a 16 GB card for H3, and 8.4 to 11.0 GB for Krea 2. A 12 GB card still trains LoRAs only; 16 GB is the fine-tune floor. Two caveats from the developer: fine-tuning is untested on AMD/ROCm so far, and Krea 2 fine-tuning realistically wants 48 GB or more of system RAM because its roughly 24 GB master copy lives in RAM.

How a 33B fine-tune fits in 16 GB

A naive full fine-tune of MiniMax H3's 33B parameters would need roughly 200 GB of memory. Fizgig rotates a trainable window through the model: every weight trains over a full cycle, but gradients and optimizer state only exist for the active slice. The frozen rest of the model is held 4-bit on the card, and the saved checkpoint is written in bf16 from a master copy that never passes through the quantizer, so the output is a clean full-precision fine-tune.

One concept explains the defaults: an epoch trains one slice of the model. The trainable window rotates each epoch, so a full cycle, typically 4 epochs, is needed for every part of the model to train once. That is why the epoch defaults look high, and why checkpoints save on cycle boundaries.

Learning rate matters more than in LoRA training

The release notes call this the one number to respect: fine-tuning wants much lower learning rates than LoRA training. Ticking Fine-tune sets a safe 1e-5 on both families. On MiniMax H3, 3e-5 is the tested faster rate and the most you should use; 1e-4 will destroy an H3 fine-tune. On Krea 2 you can experiment up to 1e-4, but results land better lower.

Checkpoint to LoRA export for ComfyUI

The built-in Checkpoint to LoRA utility diffs your fine-tune against the base model and extracts an ordinary, shareable LoRA at any rank. In the developer's testing, a rank 64 extraction was perceptually indistinguishable from the full checkpoint, in a file ComfyUI already loads normally. You can also keep the full checkpoint and set it as the family base in Preferences, then train LoRAs on top of your own fine-tuned model.