license: other
license_name: webai-non-commercial-license-ver.-1.0
license_link: LICENSE.md
base_model: webAI-Official/Air-26B-A4B-gguf
base_model_relation: finetune
library_name: llama.cpp
tags:
- gguf
- gemma4
- hierophant
- uncensored
Air 26B-A4B Hierophant 1.0 GGUF
Upstream model: webAI Air-26B-A4B GGUF
These are unofficial builds of webAI's Air 26B-A4B. The Hierophant adapter was merged into the original Air GGUF at scale 1.0. The Q6_K and IQ4_NL files are quantizations of that merge.
Uncensored variant and refusal testing
This release uses Hierophant to reduce refusal behavior. In a matched test on 17 prompts, the original Air showed an explicit refusal on all 17. The full 1.0 merge showed zero explicit refusals within the first 96 generated tokens.
| Build | Explicit refusals observed |
|---|---|
| Original Air | 17 / 17 |
| Hierophant 1.0 merge | 0 / 17 |
All 17 replies from the full 1.0 merge reached the 96-token limit, so the test does not establish whether the full replies would eventually refuse, redirect, or complete the request. Refusal behavior for Q6_K and IQ4_NL has not been measured in this test.
| File | Size | Description |
|---|---|---|
Air-26B-A4B-Hierophant-1.0.gguf |
24.96 GB | Standalone merge with mixed tensor formats |
Air-26B-A4B-Hierophant-1.0-Q6_K.gguf |
22.64 GB | Q6_K preset with shape-dependent fallbacks |
Air-26B-A4B-Hierophant-1.0-IQ4_NL.gguf |
14.02 GB | Selective IQ4_NL quantization of 30 expert down-projection tensors |
The three files loaded in llama.cpp and scored 6/9, 6/9, and 7/9 respectively on a small exact-answer benchmark at a 16,000-token context setting. The nine tasks included three long-document tasks. These results describe that sample.
License and attribution
The Air weights are distributed under the webAI Non-Commercial License ver. 1.0, which limits use to non-commercial research and education.
The base model, google/gemma-4-26B-A4B-it, is Copyright Google LLC and licensed under the Apache License 2.0. All credit for the base model goes to Google DeepMind. Gemma is a trademark of Google LLC.
Upstream quantization by webAI. The weight modification uses Hierophant.