socialtrait/flux-3-image-klein-infendpoint

Socialtrait FLUX.3 Klein deployment checkpoint

Self-contained checkpoint for the Socialtrait text-to-image service. All required model weights, tokenizer/processor files and model configurations are in this repository. Loading does not require access to another model repository.

The default A10 configuration uses the bundled Q8_0 GGUF transformer, the bundled text encoder quantized to NF4 at load time, and an FP32 VAE. It uses 16 steps, 1024 × 1024 images and seed 42, without CPU offloading or prompt upsampling. Original BF16 transformer weights are retained for explicit BF16 use on a GPU with sufficient memory. The optional upstream prompt upsampler is not included and is disabled by the supplied loader.

Contents and provenance

• transformer/flux-3-image-klein-Q8_0.gguf: unchanged Q8 checkpoint mirrored from unsloth/flux-3-image-klein-GGUF, commit 12ef13ce920d53f45fe805093910c899668b9e76.
• transformer/*.safetensors, text_encoder/, vae/: original weights retained from the existing Socialtrait checkpoint and the upstream FLUX.3 Klein model.
• checkpoint_manifest.json: original source revisions, file sizes and SHA-256 hashes for every weight file.
• handler.py, flux3_engine.py, model_config.py: HF custom handler and shared loader. Q8 and BF16 both resolve weights from the same checkpoint directory.
• modal_app.py, web.py, requirements.txt: Modal deployment and HTTP API.
• verify_checkpoint.py: offline completeness and optional checksum verification.

The packaged code and component index have been modified by Socialtrait to use the bundled assets. The original BF16 weights are unchanged; the mirrored GGUF is Unsloth's quantized derivative. Source attribution is not a runtime dependency.

HF Inference Endpoints

Select this repository and a Custom handler, with a BF16-capable CUDA GPU such as an A10. Install requirements.txt; EndpointHandler loads the mounted repository directly. Q8 is the default, with no secondary model download. Send {"inputs": "A photographic portrait"}. The response includes a base64 JPEG in image and generation details in metadata. HF manages endpoint authentication.

Modal

Download this repository's deployment files (or clone the complete repository), install modal==1.6.2, and authenticate to the intended Modal workspace. Create the socialtrait-flux3-klein secret with FLUX_API_TOKEN before running:

modal run modal_app.py --samples 1
modal deploy modal_app.py

The CPU preparation step downloads the selected weights only from this repo. The model snapshot is pinned in model_config.py; code is packaged from your checkout. See DEPLOYMENT.txt for configuration, staging CI and API details. Python dependencies and CUDA still need to be installed; this is a self-contained model checkpoint, not an operating-system/container image.

Validation

On 6 November 2026, the mounted-repository HF handler generated a valid 1024 × 1024 JPEG on NVIDIA A10 with HF_HUB_OFFLINE=1 and TRANSFORMERS_OFFLINE=1. All selected weight checksums passed before loading. One image took 15.47 seconds of generation time; peak allocated GPU memory was 12.57 GiB. This is a smoke test, not a throughput benchmark. The CPU regression suite passed 29 tests. See validation_results.json for the recorded configuration, timings and hashes of the tested deployment code.

Verify a local copy

python verify_checkpoint.py /path/to/checkpoint --hashes
python verify_checkpoint.py /path/to/checkpoint --quantization all --hashes

Run a CPU checksum check followed by one isolated A10 image generation with Hub access disabled during loading (does not deploy an HTTP endpoint):

modal run validate_deployment.py

License and attribution

Rights to use these weights are granted directly by Black Forest Labs under LICENSE.md. See NOTICE.md, the original model card in README.upstream.md, and Unsloth's preserved license, notice and model card under notices/. The text encoder's license is preserved in text_encoder/LICENSE. Packaging does not replace or alter those licenses. This is maintained by Socialtrait, not an official or endorsed Black Forest Labs product.