GLM-5.2 SIQ Fruit Instruct — BF16 source

This repository contains the serving-converted BF16 form of the assistant-masked SFT checkpoint released as malaiwah/GLM-5.2-SIQ-Fruit-Instruct. It is a 5.04B-parameter GLM-5.2 serving proxy, not the 754B GLM-5.2 model.

The source training checkpoint is fruit_v1_instruct.pt with SHA-256 32dbf82d40b88a92b8dccd563c593b5971be358cf11895eb150f18644ff93c27. The export applies the legacy training-to-serving tensor conventions, including RoPE interleave conversion and the MTP eh_proj input-half swap, while retaining all routed-expert weights in BF16.

Intended use

This snapshot is the authenticated dense source for KQuant/QSRT calibration, encoding, and matched-runtime comparisons. It is also useful for validating the Fruit checkpoint independently of the earlier SIQ/Trellis representation.

The small proxy exercises the GLM chat template, sparse DSA, 256 routed experts, and one MTP draft layer. It requires a compatible patched B12X/vLLM runtime; stock Transformers and stock vLLM do not implement this Fruit configuration.

Integrity

MANIFEST.sha256 authenticates every serving file except this model card. Verify the snapshot after download with:

sha256sum --check MANIFEST.sha256

The manifest itself has SHA-256 8f23aed5e9b12000ed103a76da772a20730ca53ab7e352d6cb94da2709165245.

Quality boundary

This is the assistant-masked SFT variant, not the QNOISE-annealed base variant. A small protocol smoke test confirmed instruction following and multilingual response behavior on the prior SIQ export; it is not a representative benchmark and no broad task-quality claim is made. Use the BF16 and QSRT forms only with matched prompts, generation settings, and runtime revisions when comparing the codec.

Provenance

Training recipe, checkpoint lineage, and review evidence are maintained in malaiwah/proxy-fruit. The base release and training data references are documented in malaiwah/GLM-5.2-SIQ-Fruit.

License

MIT, matching the packaged LICENSE file.