Distilled DeepSeek into Gemma 4 26B-A4B vs 12B. Not very useful, but I learned a lot.

So I decided to learn how to fine-tune LLMs. Read a few guides from Unsloth, poked around, then stumbled on Unsloth Studio and wanted to test it out.

I started from a set of relatively unrelated QA pairs — Natural Questions — and stripped the answers. Then I had DeepSeek v4 Pro (thinking disabled) repopulate them: - 1000 train + 200 val = 1200 requests total, cost $0.36 (~$0.0003/req). Honestly impressive on DeepSeek's side.

It's a huge pain in the butt — infested with all kinds of bugs that prevented me from using it easily. Once I figured the workflow out it was workable, but expect to debug. After that I rented a server: 2x RTX 3090, 128GB RAM, Threadripper.

Two models, to compare dense vs MoE during training: - gemma-4-26B-A4B-it-qat used both GPUs - gemma-4-12B-it-qat used one GPU

- The 26B consumed ~2x the VRAM of the 12B (28.6 vs 14.3 GB) — consistent with the MoE footprint.

- Both base models score almost identically on benchmarks, but the 26B has way more internal knowledge, which let it absorb the distillation far harder: train loss bottomed ~4x lower (0.18 vs 0.71). The eval gap was small though (1.12 vs 1.20).

- I likely overfit the 12B: eval plateaued ~1.18 around step 125–150, then drifted back up to 1.20 by step 250.

- The dense 12B was faster wall-clock (54 vs 72 min) and higher per-GPU throughput (345 vs 261 tok/s), despite the 26B using both GPUs.

I put together a dashboard image with all the hyperparameters, train/eval loss curves, grad norm, LR schedule, and timings — attached.

Models (GGUF): 1. https://huggingface.co/gwejgteheg/gemma-4-26B-A4B-it-qat-DeepSeek-distill-GGUF 2. https://huggingface.co/gwejgteheg/gemma-4-12B-IT-QAT-Q4_K_M-DeepSeek-distill-GGUF

Dataset (for reproducibility): - https://huggingface.co/datasets/gwejgteheg/natural_questions_pair/tree/main

Any feedback is appreciated and feel free to ask me any questions. Also, what kinds of fine-tunes does the community currently need?