FLUX2-TE-Trimmed7L-Research

WARNING: Experimental research artifact. Do NOT use in production.

What this is

• Model name: FLUX2-TE-Trimmed7L-Research
• Text transformer depth: 7 layers
• Vision tower depth: 24 layers (unchanged)

Caveats

• Intended for experimentation and further distillation, not end-user deployment.
• Rotary and patch positional inv_freq buffers are created at load time (not stored).

How it was built

• Shrank teacher text path to 7 layers (evenly spaced) while keeping vision depth intact.
• Stream-copied only needed tensors; other text layers were pruned.