extract-textencoder

extract-textencoder extracts the embedding and first 30 transformer layers from a compatible Hugging Face Mistral Small 3.2 24B checkpoint and packages them as a standalone FLUX.2 text encoder for ComfyUI.

This is intentionally not a universal LLM converter. The first release only supports full checkpoints derived from mistralai/Mistral-Small-3.2-24B-Instruct-2506 that retain its architecture, tokenizer, and Hugging Face tensor layout.

Environment

Use uv for environment and package management:

uv sync
uv run extract-textencoder --help

No lockfile is committed yet because this repository was prepared under a static-only constraint. Run uv lock on the machine that will perform the conversion if a reproducible environment is required.

Convert from Hugging Face

uv run extract-textencoder convert \
  --repo-id huihui-ai/Huihui-Mistral-Small-3.2-24B-Instruct-2506-abliterated \
  --revision COMMIT_HASH \
  --output mistral_3_small_flux2_huihui_bf16.safetensors

The normal Hugging Face credential sources are honored, including HF_TOKEN and credentials saved by hf auth login.

Convert a local snapshot

The directory must contain config.json, tekken.json, model.safetensors.index.json, and all referenced shard files.

uv run extract-textencoder convert \
  --local-dir /mnt/models/mistral-small-snapshot \
  --output /mnt/models/mistral_3_small_flux2_bf16.safetensors

Existing output files are protected by default. Pass --overwrite only when replacement is intended. BF16 is required by default; --allow-non-bf16 preserves floating-point source dtypes and does not quantize them.

Inspect an output

uv run extract-textencoder inspect mistral_3_small_flux2_bf16.safetensors

The inspector checks tensor names, shapes, dtypes, layer coverage, tokenizer packaging, and forbidden weights without loading or running the transformer.

For design details, see extract.md. After conversion, configure test_extracted.toml and run the standalone end-to-end test described in test_extracted.md.