Z-Image Turbo — student text encoder
Fine-tuned Qwen3-1.7B with a small adapter to imitate Qwen3-4B as Z-Image Turbo's text encoder — cutting VRAM usage while keeping quality.
I fine-tuned Qwen3-1.7B with a small adapter to imitate Qwen3-4B as the text encoder for Z-Image Turbo. The idea is simple: recreate the hidden states of Qwen3-4B with the smaller model and pass them to the DiT (Diffusion Transformer) — cutting VRAM usage while keeping quality.
Tested in fp16:
| Metric | Original (4B) | Student (1.7B) | Savings |
|---|---|---|---|
| Weight VRAM | 20.70 GB | 16.30 GB | 4.40 GB (21%) |
| Peak VRAM | 21.35 GB | 16.76 GB | 4.59 GB (22%) |
| Generation time | 3.9 s | 3.5 s | — |
I haven’t released a quantized version of this specific model yet, but existing Z-Image quants already range from ~6 GB (Q3_K_S) to ~12 GB (Q8_0), so a quantized student should be even more VRAM-efficient.