qualcomm-ai-hub-community/Qwen3-4B-Instruct-2507-Hi-Fi-fraQtl
Qwen3-4B-Instruct-2507 Hi-Fi GGUF — fraQtl
Calibration-aware GGUF builds of Qwen/Qwen3-4B-Instruct-2507: byte-matched to the leading community Q4KM with measurably lower divergence from the original weights, retrieval-verified through the full 262,144-token native context — plus a Q4_0 edition in the NPU-native fixed format.
Model Conversion Contributor: fraQtl — built by fraQtl
Model Stats:
- Quantization Type: calibration-aware per-tensor Q4KM-class (Hi-Fi) and Q4_0 (fixed format) — weights only, GGUF
- Calibration: fraQtl code+math imatrix (included in this repo)
- Maximum context length: 262,144 (native), retrieval-verified below
- Runtime: llama.cpp (CPU/GPU). GGUF works with GenieX unmodified — per the AI Hub team. (NPU/Hexagon execution not claimed — not yet verified.)
Model Download
Fidelity receipts (KLD vs the ORIGINAL model, lower is better)
Symmetric top-20 KLD vs a Q8 teacher quantized from the pinned original bf16; identical binary, template, and slices for every arm; 3 runs, drift 0:
Long-context retrieval: needle exact-match, 3 depths × 3 keys per context — 18/18 at 8K+32K (×3 runs) and 9/9 at the full 262,144-token native context (×2 runs); per-cell receipts in the canonical repo.
Q4_0 — calibration lever isolated (fixed format)
Calibration alone: −14.9% code/math, −9.2% general at identical format and source. Vs the GenieX-designated universal asset (a well-made build): we are −2.2% on code/math and +3.7% behind on general — both directions stated. K-quant-on-Hexagon NPU acceleration is unverified by us; the Q4_0 artifact is the NPU-native-format option.
Canonical repo + full receipts
Complete receipts (7 JSONs) and a runnable receipts pack (single-file runner on stock llama.cpp + teacher logit dumps — re-run every number above against any GGUF of this model): fraQtl/Qwen3-4B-Instruct-2507-Hi-Fi-GGUF
License
Apache-2.0 (inherited from the base model).
