[TroyHernandez]

FLUX 3 now ships an fp8 Qwen3-VL text encoder artifact, which lowers peak host RAM from 20.2 GB to 15.5 GB at 1024x1024 with the same speed and near-identical images.

Changes

• R/quantize_qwen3.R (new): qwen3_quantize() writes a Qwen3 or Qwen3-VL text model as an fp8 or NF4 artifact. The 252 projection linears are quantized; embeddings and norms stay bf16; the vision tower and tied LM head are dropped.
• load_qwen3_text_encoder() loads such an artifact. Its config-to-args code moved into .qwen3_args().
• download_flux3_klein(text_precision = "fp8") builds the artifact by default. flux3_load_pipeline(text_precision = "auto") uses fp8, then NF4, else bf16.
• Staging collects the fp8 weight fields for the text encoder, both whole-component and per streamed layer.
• The fp8 tensor cast is shared as .fp8_quantize().

Measured (16 GB card, 1024x1024, two renders each)

Text encoder	Host RAM peak	Time	Image vs bf16
bf16	20.2 GB	~24 s	reference
fp8	15.5 GB	~24 s	27.2 dB PSNR
NF4	13.5 GB	~25 s	15.1 dB (different composition)

• Prompt embeddings differ from bf16 by 2.7% with fp8 and about 5% with NF4, so NF4 is opt-in.
• With the 8 GB streaming plan and fp8 text, the GPU peak is 4.9 GB and the image is pixel-identical to the resident run.

Full tinytest suite: 1745 passed, 0 failed (37 new).