docs(models): add a format support matrix for every model type

[Pfannkuchensack]

Summary

Which file format loads for which model was spread over the Model Families prose and the individual family pages, and went stale whenever a format landed (Ideogram 4 and ERNIE-Image GGUF most recently). This adds one Model Format Support page under Users Guide → Models with four tables:

• Main models. Diffusers, bf16 single file, fp8 scaled, int8, nvfp4, MXFP8, GGUF, SDNQ and NF4 for all 14 families.
• Text encoders. T5, CLIP, Qwen3, Qwen3-VL (4B/8B and the truncated 32B), Qwen3.5, Qwen2.5-VL, Mistral, Ministral, Gemma-2, Gemma-4 and UMT5.
• VAEs. Single file and Diffusers, plus which family uses which VAE.
• LoRAs and adapters per family. LoRA formats, ControlNet, IP-/T2I-Adapter, Redux, PiD.

Cells are ✓ (loads), ✗ (not supported), R (refused at install with a message) or – (no such build). Model Families links to the page from its quantized-formats section. The new-model integration checklist asks to update it.

The tables reflect main as of #9708/#9709. ERNIE-Image GGUF is still ✗ here; #9725 flips that cell when it merges.

Related Issues / Discussions

Follow-up to #9708, #9709 and #9725.

QA Instructions

• pnpm -C docs build: completes.
• Every internal link on the new page resolves to a built page (16 checked against docs/dist), and the #quantized-formats anchor exists.
• The cells come from the loader registrations, the config classes and their explicit refusals in invokeai/backend/model_manager/.

Review

Docs only. No material findings.

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Meaningful regression coverage added / updated where needed; obsolete tests/code removed
• Persisted-state and API changes include required migrations / compatibility validation
• Relevant performance/efficiency opportunities considered; material claims have evidence
• Material review findings resolved and relevant checks rerun
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[joshistoast]

[P2] Mark standalone Gemma-4 files as unsupported. The matrix advertises single-file bf16 support, but Gemma-4 requires a folder containing configuration, tokenizer, and weight files. Clarify that bf16 and int8 weights require those accompanying files.

[P2] Distinguish installation refusal from loading refusal. The legend says R formats are refused during installation, but some—including ERNIE-Image quantization and Qwen3.5 cases—are rejected only when loading for generation. Remove the installation-specific wording or distinguish these stages.

[Pfannkuchensack]

Thanks, both points were right. Fixed in 73172b0.

Gemma-4: it is only recognized as a folder: config.json and the tokenizer files next to exactly one unsharded weight file, bf16 or int8. The row now reads Folder "✓ (bf16 or int8)", Single file ✗, and a note below the table says that a standalone .safetensors is not recognized.

Install vs. load: I went through every R cell against the configs and loaders. The legend now separates three outcomes:

• R: the install fails with a message. This is now only Ideogram 4 nvfp4 in the table, plus LTX-2.0/2.3 files and Q8_CR GGUFs in the notes.
• L: the model installs, then fails with a message when it is loaded for generation. This applies to the ERNIE-Image quantizations, Wan int8/nvfp4/MXFP8, LTX-2 and MiniMax H3 fp8/nvfp4, Qwen3.5, and the int8/nvfp4/fp8 encoder cells.
• ✗: the file is not recognized, so it installs as an unknown model and the reason only shows up in the server log. This covers the LTX-2 Diffusers layout, the LTX-2, MiniMax H3 and Ministral 3B GGUFs, and Wan 2.1/Animate/S2V/Fun-Control/VACE.

While checking this I also corrected the VAE note. Quantized VAEs are refused when they load for SD 3, FLUX.2, Qwen-Image, Wan/Anima and Ideogram 4. The FLUX.1 and SD 1.x/SDXL single-file VAEs do not check at all.

[joshistoast]

[P2] Include the supported FLUX.2 dev NF4 pipeline. Starter Models exposes FLUX.2 [dev] (Diffusers, NF4) from diffusers/FLUX.2-dev-bnb-4bit, and the existing family page documents it. Mark NF4 as supported inside the Diffusers folder, consistent with the Ideogram 4 row.

[Pfannkuchensack]

Right, thanks. FLUX.2 [dev]'s NF4 Diffusers pipeline (diffusers/FLUX.2-dev-bnb-4bit) loads through from_pretrained, which applies the bitsandbytes quantization_config itself. The FLUX.2 NF4 cell now reads "inside the Diffusers folder (FLUX.2 [dev])", consistent with the Ideogram 4 row. Ideogram 4 and FLUX.2 [dev] are the only families with an NF4 Diffusers starter, so no other row changes. Fixed in e8e2318.