feat: switch default hub model to FLUX.2 klein 9B (Q5_K_M GGUF)

[TimPietruskyRunPod]

Summary

Switch the default hub deployment model from FLUX.1-dev-fp8 to FLUX.2 klein 9B (Q5_K_M GGUF variant from Unsloth).

Motivation

• FLUX.2 klein 9B is a faster, higher-quality compact model released by Black Forest Labs (January 2026)
• The Unsloth Q5_K_M GGUF variant (~7GB) offers near-original quality at a fraction of the size
• Fits easily on ADA_24 (24GB) GPUs already configured for the hub (~16GB VRAM total)

Breaking Change (6.0.0)

This is a breaking change because:

• The default hub image will ship a different model
• Existing workflows using CheckpointLoaderSimple with flux1-dev-fp8.safetensors won't work on the new default image
• Users can still use the flux1-dev-fp8 image tag for backward compatibility

Files to Change

• Dockerfile - Install ComfyUI-GGUF custom node, add download block, change default MODEL_TYPE
• docker-bake.hcl - Add flux2-klein-9b target
• .github/workflows/release.yml - Add to CI matrix
• .github/workflows/manual-build-all.yml - Add build job
• .runpod/hub.json - Update description + disk size
• .runpod/README.md - Update model references + example workflow
• README.md - Add to available images list
• docs/deployment.md - Add GPU recommendation row
• test_input.json - New default workflow
• .runpod/tests_.json - New test workflow
• test_resources/workflows/workflow_flux2_klein_9b_gguf.json - New workflow file
• src/extra_model_paths.yaml - Add diffusion_models + text_encoders paths
• .changeset/flux2-klein-default.md - Major changeset

Model Details

File	Size	Directory
flux-2-klein-9b-Q5_K_M.gguf	7.02 GB	models/unet/
qwen_3_8b_fp8mixed.safetensors	8.66 GB	models/text_encoders/
flux2-vae.safetensors	336 MB	models/vae/

Total VRAM: ~16 GB on 24 GB ADA GPUs (8 GB headroom)