Downloads · 30 days
1.7K
4% of all-time downloads
EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF
gemma-4-26B-A4B-it-heretic-APEX-GGUF is a text generation model from EmanuelOverride. Use it when you need the model to write or continue text. It is set up for gguf. The card lists the license as apache-2.0.
This repository contains custom APEX (Adaptive Precision for EXpert Models) -inspired quants, tuned specifically to the underlying mixture-of-experts architecture and focusing on intermediary-size tiers.
Downloads · 30 days
1.7K
4% of all-time downloads
All-time downloads
47.8K
Public
Repo size
258 GB
Likes
6
Public
Click a slice to open those files.
.gguf209 GB · 100%
From the Hugging Face model README
This repository contains custom APEX (Adaptive Precision for EXpert Models) -inspired quants, tuned specifically to the underlying mixture-of-experts architecture and focusing on intermediary-size tiers.
⚠️ Disclaimer: The layouts hosted here are custom-crafted and experimental. If you are looking for the standard, stable, and predictable APEX suite of this model, please visit mudler's excellent repository here: (https://huggingface.co/mudler/gemma-4-26B-A4B-it-heretic-APEX-GGUF), where a complete baseline, uniform APEX suite is maintained. All credit for the underlying architecture, abliteration mechanics, and original model weights belongs entirely to the various upstream authors, and the imatrix-dataset was generated by mradermacher.
Why APEX? Regular uniform quantization applies compression equally across all layers, which can degrade MoE performance. The APEX layout locks the routing blocks at high-precision and shields vital shared experts, shifting aggressive compression strictly to redundant mid-layer tensors.
Why custom? The standard APEX configurations generally lack an intermediary tier, between high compression Mini & Compact and high precision Balanced & Quality. In trying to create such an intermediary tier, fairly subtle adjustments to the mixed precision distributions can have an outsized influence on model coherence and behavior, and exploring this narrow range in the open, could hopefully result in the creation of highly usable, high-performance layer configurations.
In my initial attempts at creating an intermediary size-tier between Compact and Balanced, I scaled up the basic logic of the Compact configuration by one quant level:
14,600.94 MiB (4.85 BPW) — Performance King on legacy hardware. Employs a [Q5_K -> Q4_K -> Q3_K] stepped gradient across the mid-layer experts while shielding context gates.14,790.00 MiB (4.92 BPW) — Flat Q4_K expert mapping across all up/down tensors. Known and stable size-to-precision efficiency curve.12,899.38 MiB (4.29 BPW) — Aggressive Q4_K to Q3_K mid-expert trimming. Extensive use of Q3_K and how the ratio influences lookup tables probably causes it to balloon in memory? My best guess. Probably non-viable econ solution and won't be done in the future.Following on from the Harmonic Resonance branch, the Harmonic layouts further optimize layer contiguity for performance, by following strict precision tiers across all 30 layers (High/Medium/Low), while priorizing high-precision tiers for Global Attention (GA) layers, preventing aggressive compression in order to anchor deep-context tracking.
14,369.74 MiB (4.78 BPW) — High-performance configuration. Condenses the architecture into 3 strict tiers (5/15/10). Shifts high-precision blocks to coincide with global attention pulses. In testing: Faster than Upscale.Premium, with better long-context detail attention.15,804.70 MiB (5.25 BPW) — Medium-High-precision/Performance configuration. Eliminates the Q3_K compression valley entirely to enforce a flat, 2-tier matrix (10/20). In testing: Replaces and outperforms Upscale.Balanced at slightly larger overhead, while shielding long-context detail attention.The Mindful layouts build on the Harmonic framework, balancing precision-tier distributions while implementing additional buffers around the GA-layers.
15,606.86 MiB (5.19 BPW) — Experimental Semantic Anchor. Follows 15/9/6 distribution. Trades uniform end-precision for comprehensive buffering of all Global Attention layers. Expectations: Slight edge in context stability and logical coherence under heavy prompt load.15,587.68 MiB (5.18 BPW) — Experimental Fluidity Engine. Built on a 16/7/7 topology. Trades early intermediate buffers for an uncompromised, clean high-precision runway across the final formatting layers. Expectations: Slight edge in inference speed and final execution.While the initial i-Upscale suite succeeded in creating at least two highly efficient, streamlined intermediary sizes, it still remained relatively close to the BPW baseline of the APEX Compact. Echo Valley steps away from flat segmentation to explore a more distinct Neural Topology—treating the model's 30-layer transformer brain as a dynamic, continuous acoustic waveform, attempting to shape the network into a bowl or valley as seen from the side, upscaling the ridges even further while expanding the scope of the compressed middle.
i-EchoValley.Plus: 16,952.88 MiB (5.64 BPW) — (6-2-2-10-2-2-6 Split).
i-EchoValley.Minus: 14,752.19 MiB (4.90 BPW) — (6-3-12-3-6 Split).
i-EchoValley.Inverse: 16,257.13 MiB (5.40 BPW) — (4-3-3-10-3-3-4 Split).
EDIT: After further testing and consideration, segmentation of ramps seems to primarily have a slightly negative impact on performance by introducing an additional tier/ratio; any other influence too subtle for me to detect. Overall concept still useful, Plus/Minus varieties can be optimized further to serve Quality/Economy roles.
An advanced piece of custom neural topology synchronized to the model's native 6-layer global attention pulse relays (Layers 5, 11, 17, 23, 29). By eliminating the prime-number noise of 5-bit layers and shifting non-divisible legacy fallback rows with calibrated Importance Quantization (IQ4_NL), the data wave moves in a continuous binary octave sequence (Q8_0 -> Q6_K -> Q4_K -> Q3_K). From testing: Elevated lateral thinking, crystalline prose focus, and higher intentionality during long-context interactions, but at the cost of some performance and memory overhead.
15,582.23 MiB (5.18 BPW)EDIT: After further testing and consideration, there's a lot to learn from this one. NL format not particularly worth it, at least on my hardware, making a Q5-"level" quant devoid of 5-bit layers quite reasonable upgrade, but needs further research, and reinforcing GA layers arguably more impactful than trying to preserve edge-layer mixed experts, at least for long-context detail attention. Underlying design can be optimized for further performance, by aligning individual blocks to layer-precision tiers, eliminating niche tiers for greater contiguity across entire structure.
Please open a thread, share your experience, and post your impressions.
| Parameter | Value |
|---|---|
| start_layer_index | 10 |
| end_layer_index | 30 |
| preserve_good_behavior_weight | 0.5480 |
| steer_bad_behavior_weight | 0.0009 |
| overcorrect_relative_weight | 0.5868 |
| neighbor_count | 14 |
| Metric | This model | Original model (google/gemma-4-26B-A4B-it) |
|---|---|---|
| KL divergence | 0.0499 | 0 (by definition) |
| Refusals | 11/100 | 100/100 |
Gemma is a family of open models built by Google DeepMind. Gemma 4 models are multimodal, handling text and image input (with audio supported on small models) and generating text output. This release includes open-weights models in both pre-trained and instruction-tuned variants. Gemma 4 features a context window of up to 256K tokens and maintains multilingual support in over 140 languages.
Featuring both Dense and Mixture-of-Experts (MoE) architectures, Gemma 4 is well-suited for tasks like text generation, coding, and reasoning. The models are available in four distinct sizes: E2B, E4B, 26B A4B, and 31B. Their diverse sizes make them deployable in environments ranging from high-end phones to laptops and servers, democratizing access to state-of-the-art AI.
Gemma 4 introduces key capability and architectural advancements:
Reasoning – All models in the family are designed as highly capable reasoners, with configurable thinking modes.
Extended Multimodalities – Processes Text, Image with variable aspect ratio and resolution support (all models), Video, and Audio (featured natively on the E2B and E4B models).
Diverse & Efficient Architectures – Offers Dense and Mixture-of-Experts (MoE) variants of different sizes for scalable deployment.
Optimized for On-Device – Smaller models are specifically designed for efficient local execution on laptops and mobile devices.
Increased Context Window – The small models feature a 128K context window, while the medium models support 256K.
Enhanced Coding & Agentic Capabilities – Achieves notable improvements in coding benchmarks alongside native function-calling support, powering highly capable autonomous agents.
Native System Prompt Support – Gemma 4 introduces native support for the system role, enabling more structured and controllable conversations.
Gemma 4 models are designed to deliver frontier-level performance at each size, targeting deployment scenarios from mobile and edge devices (E2B, E4B) to consumer GPUs and workstations (26B A4B, 31B). They are well-suited for reasoning, agentic workflows, coding, and multimodal understanding.
The models employ a hybrid attention mechanism that interleaves local sliding window attention with full global attention, ensuring the final layer is always global. This hybrid design delivers the processing speed and low memory footprint of a lightweight model without sacrificing the deep awareness required for complex, long-context tasks. To optimize memory for long contexts, global layers feature unified Keys and Values, and apply Proportional RoPE (p-RoPE).
| Property | E2B | E4B | 31B Dense |
|---|---|---|---|
| Total Parameters | 2.3B effective (5.1B with embeddings) | 4.5B effective (8B with embeddings) | 30.7B |
| Layers | 35 | 42 | 60 |
| Sliding Window | 512 tokens | 512 tokens | 1024 tokens |
| Context Length | 128K tokens | 128K tokens | 256K tokens |
| Vocabulary Size | 262K | 262K | 262K |
| Supported Modalities | Text, Image, Audio | Text, Image, Audio | Text, Image |
| Vision Encoder Parameters | ~150M | ~150M | ~550M |
| Audio Encoder Parameters | ~300M | ~300M | No Audio |
The "E" in E2B and E4B stands for "effective" parameters. The smaller models incorporate Per-Layer Embeddings (PLE) to maximize parameter efficiency in on-device deployments. Rather than adding more layers or parameters to the model, PLE gives each decoder layer its own small embedding for every token. These embedding tables are large but are only used for quick lookups, which is why the effective parameter count is much smaller than the total.
| Property | 26B A4B MoE |
|---|---|
| Total Parameters | 25.2B |
| Active Parameters | 3.8B |
| Layers | 30 |
| Sliding Window | 1024 tokens |
| Context Length | 256K tokens |
| Vocabulary Size | 262K |
| Expert Count | 8 active / 128 total and 1 shared |
| Supported Modalities | Text, Image |
| Vision Encoder Parameters | ~550M |
The "A" in 26B A4B stands for "active parameters" in contrast to the total number of parameters the model contains. By only activating a 4B subset of parameters during inference, the Mixture-of-Experts model runs much faster than its 26B total might suggest. This makes it an excellent choice for fast inference compared to the dense 31B model since it runs almost as fast as a 4B-parameter model.
These models were evaluated against a large collection of different datasets and metrics to cover different aspects of text generation. Evaluation results marked in the table are for instruction-tuned models.
| Gemma 4 31B | Gemma 4 26B A4B | Gemma 4 E4B | Gemma 4 E2B | Gemma 3 27B (no think) | |
|---|---|---|---|---|---|
| MMLU Pro | 85.2% | 82.6% | 69.4% | 60.0% | 67.6% |
| AIME 2026 no tools | 89.2% | 88.3% | 42.5% | 37.5% | 20.8% |
| LiveCodeBench v6 | 80.0% | 77.1% | 52.0% | 44.0% | 29.1% |
| Codeforces ELO | 2150 | 1718 | 940 | 633 | 110 |
| GPQA Diamond | 84.3% | 82.3% | 58.6% | 43.4% | 42.4% |
| Tau2 (average over 3) | 76.9% | 68.2% | 42.2% | 24.5% | 16.2% |
| HLE no tools | 19.5% | 8.7% | - | - | - |
| HLE with search | 26.5% | 17.2% | - | - | - |
| BigBench Extra Hard | 74.4% | 64.8% | 33.1% | 21.9% | 19.3% |
| MMMLU | 88.4% | 86.3% | 76.6% | 67.4% | 70.7% |
| Vision | |||||
| MMMU Pro | 76.9% | 73.8% | 52.6% | 44.2% | 49.7% |
| OmniDocBench 1.5 (average edit distance, lower is better) | 0.131 | 0.149 | 0.181 | 0.290 | 0.365 |
| MATH-Vision | 85.6% | 82.4% | 59.5% | 52.4% | 46.0% |
| MedXPertQA MM | 61.3% | 58.1% | 28.7% | 23.5% | - |
| Audio | |||||
| CoVoST | - | - | 35.54 | 33.47 | - |
| FLEURS (lower is better) | - | - | 0.08 | 0.09 | - |
| Long Context | |||||
| MRCR v2 8 needle 128k (average) | 66.4% | 44.1% | 25.4% | 19.1% | 13.5% |
Gemma 4 models handle a broad range of tasks across text, vision, and audio. Key capabilities include:
You can use all Gemma 4 models with the latest version of Transformers. To get started, install the necessary dependencies in your environment:
pip install -U transformers torch accelerate
Once you have everything installed, you can proceed to load the model with the code below:
from transformers import AutoProcessor, AutoModelForCausalLM
MODEL_ID = "google/gemma-4-26B-A4B-it"
# Load model
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForCausalLM.from_pretrained(
MODEL_ID,
dtype="auto",
device_map="auto"
)
Once the model is loaded, you can start generating output:
# Prompt
messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Write a short joke about saving RAM."},
]
# Process input
text = processor.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=True,
enable_thinking=False
)
inputs = processor(text=text, return_tensors="pt").to(model.device)
input_len = inputs["input_ids"].shape[-1]
# Generate output
outputs = model.generate(**inputs, max_new_tokens=1024)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
# Parse output
processor.parse_response(response)
To enable reasoning, set enable_thinking=True and the parse_response function will take care of parsing the thinking output.
Below, you will also find snippets for processing audio (E2B and E4B only), images, and video alongside text:
<details> <summary>Code for processing Audio</summary>Instead of using AutoModelForCausalLM, you can use AutoModelForMultimodalLM to process audio. To use it, make sure to install the following packages:
pip install -U transformers torch librosa accelerate
You can then load the model with the code below:
from transformers import AutoProcessor, AutoModelForMultimodalLM
MODEL_ID = "google/gemma-4-E2B-it"
# Load model
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
MODEL_ID,
dtype="auto",
device_map="auto"
)
Once the model is loaded, you can start generating output by directly referencing the audio URL in the prompt:
# Prompt - add audio before text
messages = [
{
"role": "user",
"content": [
{"type": "audio", "audio": "https://raw.githubusercontent.com/google-gemma/cookbook/refs/heads/main/Demos/sample-data/journal1.wav"},
{"type": "text", "text": "Transcribe the following speech segment in its original language. Follow these specific instructions for formatting the answer:\n* Only output the transcription, with no newlines.\n* When transcribing numbers, write the digits, i.e. write 1.7 and not one point seven, and write 3 instead of three."},
]
}
]
# Process input
inputs = processor.apply_chat_template(
messages,
tokenize=True,
return_dict=True,
return_tensors="pt",
add_generation_prompt=True,
).to(model.device)
input_len = inputs["input_ids"].shape[-1]
# Generate output
outputs = model.generate(**inputs, max_new_tokens=512)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
# Parse output
processor.parse_response(response)
</details>
<details>
<summary>Code for processing Images</summary>
Instead of using AutoModelForCausalLM, you can use AutoModelForMultimodalLM to process images. To use it, make sure to install the following packages:
pip install -U transformers torch torchvision accelerate
You can then load the model with the code below:
from transformers import AutoProcessor, AutoModelForMultimodalLM
MODEL_ID = "google/gemma-4-26B-A4B-it"
# Load model
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
MODEL_ID,
dtype="auto",
device_map="auto"
)
Once the model is loaded, you can start generating output by directly referencing the image URL in the prompt:
# Prompt - add image before text
messages = [
{
"role": "user", "content": [
{"type": "image", "url": "https://raw.githubusercontent.com/google-gemma/cookbook/refs/heads/main/Demos/sample-data/GoldenGate.png"},
{"type": "text", "text": "What is shown in this image?"}
]
}
]
# Process input
inputs = processor.apply_chat_template(
messages,
tokenize=True,
return_dict=True,
return_tensors="pt",
add_generation_prompt=True,
).to(model.device)
input_len = inputs["input_ids"].shape[-1]
# Generate output
outputs = model.generate(**inputs, max_new_tokens=512)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
# Parse output
processor.parse_response(response)
</details>
<details>
<summary>Code for processing Videos</summary>
Instead of using AutoModelForCausalLM, you can use AutoModelForMultimodalLM to process videos. To use it, make sure to install the following packages:
pip install -U transformers torch torchvision torchcodec librosa accelerate
You can then load the model with the code below:
from transformers import AutoProcessor, AutoModelForMultimodalLM
MODEL_ID = "google/gemma-4-26B-A4B-it"
# Load model
processor = AutoProcessor.from_pretrained(MODEL_ID)
model = AutoModelForMultimodalLM.from_pretrained(
MODEL_ID,
dtype="auto",
device_map="auto"
)
Once the model is loaded, you can start generating output by directly referencing the video URL in the prompt:
# Prompt - add video before text
messages = [
{
'role': 'user',
'content': [
{"type": "video", "video": "https://github.com/bebechien/gemma/raw/refs/heads/main/videos/ForBiggerBlazes.mp4"},
{'type': 'text', 'text': 'Describe this video.'}
]
}
]
# Process input
inputs = processor.apply_chat_template(
messages,
tokenize=True,
return_dict=True,
return_tensors="pt",
add_generation_prompt=True,
).to(model.device)
input_len = inputs["input_ids"].shape[-1]
# Generate output
outputs = model.generate(**inputs, max_new_tokens=512)
response = processor.decode(outputs[0][input_len:], skip_special_tokens=False)
# Parse output
processor.parse_response(response)
</details>
For the best performance, use these configurations and best practices:
Use the following standardized sampling configuration across all use cases:
temperature=1.0top_p=0.95top_k=64Compared to Gemma 3, the models use standard system, assistant, and user roles. To properly manage the thinking process, use the following control tokens:
<|think|> token at the start of the system prompt. To disable thinking, remove the token.<|channel>thought\n[Internal reasoning]<channel|><|channel>thought\n<channel|>[Final answer][!Note] Note that many libraries like Transformers and llama.cpp handle the complexities of the chat template for you.
Aside from variable aspect ratios, Gemma 4 supports variable image resolution through a configurable visual token budget, which controls how many tokens are used to represent an image. A higher token budget preserves more visual detail at the cost of additional compute, while a lower budget enables faster inference for tasks that don't require fine-grained understanding.
Use the following prompt structures for audio processing:
Transcribe the following speech segment in {LANGUAGE} into {LANGUAGE} text.
Follow these specific instructions for formatting the answer:
* Only output the transcription, with no newlines.
* When transcribing numbers, write the digits, i.e. write 1.7 and not one point seven, and write 3 instead of three.
Transcribe the following speech segment in {SOURCE_LANGUAGE}, then translate it into {TARGET_LANGUAGE}.
When formatting the answer, first output the transcription in {SOURCE_LANGUAGE}, then one newline, then output the string '{TARGET_LANGUAGE}: ', then the translation in {TARGET_LANGUAGE}.
All models support image inputs and can process videos as frames whereas the E2B and E4B models also support audio inputs. Audio supports a maximum length of 30 seconds. Video supports a maximum of 60 seconds assuming the images are processed at one frame per second.
Data used for model training and how the data was processed.
Our pre-training dataset is a large-scale, diverse collection of data encompassing a wide range of domains and modalities, which includes web documents, code, images, audio, with a cutoff date of January 2025. Here are the key components:
The combination of these diverse data sources is crucial for training a powerful multimodal model that can handle a wide variety of different tasks and data formats.
Here are the key data cleaning and filtering methods applied to the training data:
As open models become central to enterprise infrastructure, provenance and security are paramount. Developed by Google DeepMind, Gemma 4 undergoes the same rigorous safety evaluations as our proprietary Gemini models.
Gemma 4 models were developed in partnership with internal safety and responsible AI teams. A range of automated as well as human evaluations were conducted to help improve model safety. These evaluations align with Google’s AI principles, as well as safety policies, which aim to prevent our generative AI models from generating harmful content, including:
For all areas of safety testing, we saw major improvements in all categories of content safety relative to previous Gemma models. Overall, Gemma 4 models significantly outperform Gemma 3 and 3n models in improving safety, while keeping unjustified refusals low. All testing was conducted without safety filters to evaluate the model capabilities and behaviors. For both text-to-text and image-to-text, and across all model sizes, the model produced minimal policy violations, and showed significant improvements over previous Gemma models' performance.
These models have certain limitations that users should be aware of.
Multimodal models (capable of processing vision, language, and/or audio) have a wide range of applications across various industries and domains. The following list of potential uses is not comprehensive. The purpose of this list is to provide contextual information about the possible use-cases that the model creators considered as part of model training and development.
The development of vision-language models (VLMs) raises several ethical concerns. In creating an open model, we have carefully considered the following:
Risks identified and mitigations:
At the time of release, this family of models provides high-performance open vision-language model implementations designed from the ground up for responsible AI development compared to similarly sized models.