Downloads · 30 days
0
HUSRCF/gemma-4-26b-a4b-it-hfq4g128-tailfix
gemma-4-26b-a4b-it-hfq4g128-tailfix is a machine learning model from HUSRCF. Use it for the machine learning task on the model card, and read the license before you ship it in a product. It is set up for hipfire. The card lists the license as gemma.
This repository contains a text-only Hipfire HFQ artifact converted locally from the Google Gemma 4 26B-A4B-it BF16 checkpoint. It is not an official Google quantization and was not produced or released by the Hipfire…
Downloads · 30 days
0
Access
Public
Updated Sep 7, 2026
Repo size
15.6 GB
Likes
0
Public
Click a slice to open those files.
.hf415.6 GB · 100%
From the Hugging Face model README
This repository contains a text-only Hipfire HFQ artifact converted locally from the Google Gemma 4 26B-A4B-it BF16 checkpoint. It is not an official Google quantization and was not produced or released by the Hipfire maintainers.
google/gemma-4-26B-A4B-ithttps://modelscope.cn/models/google/gemma-4-26B-A4B-itconfig.json SHA-256: ed0c1eb3633de771906e9ba004a44cc5635bcc06ee2062077c3d2e88a50707d3model.safetensors.index.json SHA-256: 907826a6e46ff454272bd6db1fee629d5531a2303be22986d825a0871d7dc7a7hipfire-quantize 0.3.0 at Hipfire commit 27f333ff1HIPFIRE_GIT_COMMIT=27f333ff1 hipfire-quantize \
--input /path/to/google-gemma-4-26B-A4B-it \
--output gemma-4-26b-a4b-it.hfq4g128-tailfix.hf4 \
--format hf4 \
--arch-id 13 \
--source-url https://modelscope.cn/models/google/gemma-4-26B-A4B-it \
--license LicenseRef-Gemma
The output embeds this source URL, license reference, format, tool version, and converter commit in hipfire_provenance.
gemma-4-26b-a4b-it.hfq4g128-tailfix.hf415,632,064,572 bytesccd0eb6881d170829a1d9d092b73a0956025548c34fbad3438f588488f0b263413 (gemma4)8,277Deterministic recipe census:
| Wire type | Tensor records |
|---|---|
| F16 | 361 |
| Q8F16 | 147 |
| HFQ6-G256 | 549 |
| HFQ4-G256 | 3,380 |
| HFQ4-G128 | 3,840 |
The G128 records preserve ragged expert rows that cannot be represented safely as complete G256 groups. The quantizer reported mean quantization error 0.00003919 and maximum quantization error 0.04241335.
The vision/auxiliary tensors were not included: the converter reported 572,794,416 skipped parameters. This artifact is intended for Hipfire's text-generation Gemma4 lowered/MoE path, not multimodal inference.
The weight payload is byte-identical to the earlier local tail-safe artifact used for Hipfire PR #667 validation; only the HFQ metadata region changed to add provenance. That payload passed Hipfire Redline HIP/PM4 parity and user-facing serve batteries on gfx1100 (Radeon Pro W7900) and gfx1201 (Radeon AI Pro R9700). Those results are commit- and runtime-specific evidence, not a general quality or performance guarantee.
Use of this artifact remains subject to the Gemma license and terms associated with the base model.