SummaryGemma 4 26B A4B is Google DeepMind's mixture-of-experts model with 25.2B total and 3.8B active parameters, 256K context and image input, released under Apache 2.0. Its refusal-reduced variants include huihui-ai's abliterated weights, an abliterated QAT build, two Heretic builds that publish refusal and KL figures, and a HauhauCS GGUF with an undisclosed method.
What the base model is
Gemma 4 26B A4B is the medium mixture-of-experts model in Google DeepMind’s Gemma 4 family. The “A4B” means about 4 billion parameters are active per token; the card gives 25.2B total and 3.8B active, with 128 experts plus one shared expert and 8 used per token. It takes text and images and supports a configurable thinking mode.
Two details matter for running it. First, low active parameters make it fast per token, but every expert must sit in memory, so it needs about as much memory as a 25B dense model. Second, its attention is hybrid: 25 of its 30 layers use a sliding window of 1,024 tokens and only 5 keep the full context. That keeps the KV cache small even at long context, which the VRAM estimator models from the configuration file.
Google publishes Gemma 4 under the Apache License 2.0. Earlier Gemma generations used the Gemma Terms of Use, and some variant repos still carry that older tag; the base license is the one that governs the weights.
The variants at a glance
| Variant |
Method (per card) |
What the card reports |
| huihui-ai abliterated |
Abliteration |
First 5 layers and the experts not edited; may still add warnings but not refuse |
| huihui-ai QAT abliterated (GGUF) |
Abliteration of Google’s quantization-aware checkpoint |
No measurements |
| coder3101 heretic |
Heretic 1.2.0, ARA method |
11/100 refusals vs 100/100; KL divergence 0.0499 |
| llmfan46 ultra-uncensored heretic |
Heretic 1.2.0, ARA method |
3/100 refusals vs 100/100; KL divergence 0.1237 |
| HauhauCS QAT Balanced (GGUF) |
Not disclosed |
“0/465 refusals” on an unpublished test; MTP draft head |
The two Heretic builds show the usual tradeoff. The llmfan46 build reports fewer refusals than coder3101’s but a KL divergence about two and a half times higher, meaning its predictions moved further from the original. Both numbers come from Heretic’s default prompt set as run by each creator, so treat them as a comparison between these two builds of the same base, not as a quality score.
Quantization-aware training (QAT) means Google trained a checkpoint to lose less quality when compressed to 4 bits. Two variants start from that checkpoint, which is why HauhauCS ships only a Q4_K_M file and says higher-precision files would add size with little gain.
Which one for which need
- For published numbers and two hosting options: coder3101’s Heretic build. Venice serves it in a trusted execution environment with end-to-end encryption, according to its model list, with a 64,000-token context and 4,096 tokens of output, but lists neither image input nor function calling for that offering. Featherless serves the same repo with a 32,768-token context.
- To run at 4-bit on one consumer GPU: one of the QAT-based GGUF files. Weights take roughly 15 GiB at Q4_K_M, plus a 1.2 GB vision projector if you need image input.
- For the most documentation on what changed: huihui-ai’s release, which says which layers and modules it left alone.
MolochAI lists gemma4-26b-a4b-qat, described as a refusal-reduced QAT build, without naming the weights. Rates for all of these are below.
Limits and open questions
The huihui-ai card says it left “the 256 experts per layer” unedited, while Google’s card lists 128 experts plus one shared expert. It is probably a typo, but it is the kind of mismatch to note when a card is your only documentation. The HauhauCS repo is tagged with the Gemma Terms of Use license although Gemma 4 is Apache 2.0.
No variant publishes a quality benchmark after the edit. Before relying on image input, test it: editing the language layers can change how the model describes what it sees.
Ordered as catalogued. Method, layers and claims are as each creator states them; Atlas has not tested any of these variants.
Estimate for Gemma 4 26B A4B (instruction-tuned) with 32,768 tokens of context, one sequence, FP16 KV cache and 10% overhead. Variants of the same architecture need about the same.