RESMP-DEV/Qwen3.8-Flash-Next-GPTQ-MXFP4

Qwen3.8-Flash-Next calibrated GPTQ MXFP4 research archive

This is a campaign-owned MLX research conversion of Qwen/Qwen3.8-Flash-Next. The resident transformer backbone is native MXFP4 with group size 32. The n-gram PLE values, embedding, and language-model head remain BF16.

The archive preserves the packed backbone and BF16 PLE source planes together. It is intended for reproducibility and continued MLX runtime research; it is not a drop-in stock MLX-LM checkpoint.

Quantization and coverage

Property	Value
Source revision	de4b8e4d43b917e7706784d8bb445c9af86a3540
Backbone weight format	MXFP4 E2M1
Scale format	E8M0
Group size	32
Stage A tensor records	74,400
Layers completed	48/48
Packed tensors	2,160
Safetensors shards	34
Packed model bytes	171,213,918,824
PLE, embedding, and head	BF16

The GPTQ calibration used the same private pinned activation corpus and sequential native-scale solver as the Qwen3.8-27B MXFP4 release. Full PLE planes were not materialized in process during packing.

Runtime boundary

Stock MLX-LM does not load this archive directly because the large PLE planes remain in source shards and require a PLE-aware selected-row reader. The research runtime parses safetensors byte ranges, fetches only requested rows, deduplicates requests, coalesces contiguous ranges, and keeps a bounded row cache.

The loader-side research artifact contains 2,032 backbone keys and 68,813,407,880 model bytes after excluding the PLE-only planes. It is not standalone either: it must be paired with the PLE shards in this archive or an exact copy of their rows.

Text-only MRoPE parity was checked against the installed Transformers reference. The maximum cosine and sine differences were approximately 0.00195 and 0.00189, attributable to BF16 rounding. This does not validate vision inputs.

Verification and limitations

• All Stage A tensor records were consumed.
• All 34 packed shards were readable and SHA-256 checked.
• The safetensors index maps every packed tensor exactly.
• Two real fixed-prompt Apple Silicon runs generated 32/32 tokens with byte-identical text.
• Those runs used 46.80 GB and 47.25 GB peak process RSS.
• PLE gather time was 29.03 ms and 27.29 ms cumulative over 32 generated tokens; storage access was not the measured bottleneck.

These checks prove bounded loading and deterministic text on one prompt. They do not establish perplexity, general generation quality, multimodal parity, or production throughput. The custom PLE-aware MLX-LM reader is not included in this model repository.