Mistral-Large-4-EXL3-K4-Core-K6-HQA

HQA = High Quality Attention. This is wrldsuksgo2mars/Mistral-Large-4-EXL3-K4-Core-K6 (pinned at 03c80aebaa483d55760703ae5bc0c9734c9d67d6) with its main attention projections restored to the original release's own weights: native F8_E4M3 from mistralai/Mistral-Large-4-1T-A52B at f12dabdc0c1e99e7f7924461a6f308859b05af4d, copied byte-for-byte together with their original FP32 block-128 scales (no casting, requantization, padding or transposition). The expert quantization and every other tensor are the parent's, unchanged. HQA describes the format and fidelity of these tensors; it is not a measured benchmark claim.

Restored tensors

144 projections in total. The compact view keeps the parent's quantization of these modules exactly, including the lower-bit MTP attention where the parent used it. Indexer projections, draft weights that were already native, and all other modules are untouched.

Two attention views

This repository holds two complete, mutually exclusive views of the same files:

The default view maps each restored projection to its native tensors in the hqa-attention-* shards and omits that projection's EXL3 buffers. The compact-attention index is the parent's index byte-for-byte: it maps the original EXL3 attention buffers (which remain in the inherited shards, deliberately, for this view) and never references the hqa-attention-* shards, so a compact download can skip them. Its config and quantization metadata are the parent's, with only sibling-file references redirected to the compact filenames. These alternate filenames are a cuteafd convention, not a Hugging Face mode switch: generic loaders read the default files only, and a valid safetensors index does not imply generic Transformers/EXL3 support.

Loader support. A loader must select one bundle (index, config and quantization metadata together) for the lifetime of the loaded model, treat the selected index as authoritative, and refuse a missing or mixed bundle. The cuteafd --compact-attention option that selects the compact bundle is future work and is not implemented yet; per-family cuteafd loader support for the HQA view lands separately. Until then this checkpoint is published for that work; do not assume a current cuteafd release serves either view of it.

Sizes

The views share almost all bytes; do not add them. New shards (balanced across 4, spread 0 bytes):

Validation performed

• Every replacement tensor was compared byte-for-byte with its pinned source extent (7,249,526,784 bytes); shapes, dtypes, scale pairing and shard hashes checked.
• Both views were opened through an explicit selection harness implementing the contract above: every indexed tensor resolves in its file header; each view's payload equals its total_size; no projection exposes both native and EXL3 storage; the compact view equals the pinned parent's tensor selection and payload exactly and opens none of the native shards (and loads with them physically absent); the views have distinct identities.
• Inherited files are server-side copies of the pinned parent. Publication completes only after the remote tree matches the planned inventory exactly: inherited LFS SHA256 values equal the pinned parent's, and new files equal the local write records.
• The cuteafd planner (cuteafd plan at cuteafd 3ba032b62) on a directory exposing only the compact bundle under the canonical filenames, with the native shards absent, reproduced the pinned parent plan exactly. The same planner does not yet execute the HQA bundle: GLM placement: GLM layer 0 attention mixes native and EXL3 storage; complete operator groups required; routed_expert: 1 of 27264 weights not executable, e.g. routed experts (read_expert_catalog): index/header tensor count mismatch in model-00001-of-00073.safetensors. That is the separate family loader work; the artifact contract does not depend on it.
• Not performed: whole-model generation, quality or performance evaluation of the HQA view.

Provenance

hqa-provenance.json binds both views, the parent and source revisions, every restored tensor's source byte extent, shard membership and hashes, and the producing tool. README.parent.md, quantization-provenance.json and quantization-errors/ are the parent's records: they describe the compact-attention view and every unchanged module, not the restored native modules. License and usage files are inherited unchanged from the parent.

Produced by quantization/hqa.py in cuteafd (CPU and I/O only).