cyberneurova-gemma-4-26B-A4B-GGUF
Quantized GGUF distribution of cyberneurova-gemma-4-26B-A4B — CyberNeurova's cracked build of the Gemma-4 26B A4B MoE instruction model.
Part of the CyberNeurova Cracked Gemma Series — a curated set of Gemma-family models with refusal behavior removed at the weight level, standardized under the CyberNeurova brand and shipped as production-ready GGUF quants for self-hosted deployment.
What "cracked" means here
These are abliterated Gemma-4 models. Abliteration identifies the direction in the model's residual stream that corresponds to refusal behavior and projects it out of the weights. The model retains its original reasoning, knowledge, and capabilities but no longer produces default refusal responses. It is not retrained, not jailbroken, not prompt-hacked — the refusal circuit itself is neutralized in the weights.
We ship these under the CyberNeurova namespace for:
- Consistent naming across our served model catalogue
- Standardized imatrix-calibrated GGUF quantization
- Unified configuration, split layout, and chat-template handling
- Direct integration with CyberNeurova inference infrastructure
Model details
| Attribute | Value |
|---|---|
| Architecture | gemma4 (Gemma-4 MoE) |
| Total parameters | 26B |
| Active parameters | 4B |
| Modalities | Text, Image (vision-capable), Tool use, Thinking mode |
| Base model | google/gemma-4-26B-A4B |
| License | Gemma Terms of Use |
Available quantizations
All quants are imatrix-calibrated — strongly recommended for MoE models to preserve expert routing fidelity at low bit widths.
| Quant | Bits | Use case |
|---|---|---|
Q2_K | 2 | Smallest footprint. Casual chat; noticeable quality drop on reasoning. |
IQ3_XS | 3 | Smart 3-bit. Better than Q3_K_S at similar size. |
Q3_K_S | 3 | Small, reasonable quality. |
IQ3_M | 3 | Best 3-bit quality/size tradeoff. |
Q3_K_M | 3 | Standard 3-bit. |
Q3_K_L | 3 | Larger 3-bit, closer to Q4. |
IQ4_XS | 4 | Smart 4-bit, smaller than Q4_K_S with near-equivalent quality. |
Q4_0 | 4 | Legacy format, broad compatibility. |
IQ4_NL | 4 | Non-linear 4-bit, good for ARM/CPU. |
Q4_K_S | 4 | Small 4-bit. |
Q4_K_M | 4 | Recommended default. Best size/quality ratio for most workloads. |
Q5_K_S | 5 | Small 5-bit. |
Q5_K_M | 5 | Higher quality 5-bit. |
Q6_K | 6 | Near-lossless. |
Q8_0 | 8 | Effectively lossless. Recommended for serving if VRAM allows. |
MXFP4_MOE | 4 | Expert weights in MXFP4, other tensors in Q8_0. Optimized for Gemma-4's MoE expert structure. |
Choosing a quant
- Chat / general use:
Q4_K_MorMXFP4_MOE - Coding / tool calling:
Q5_K_MorQ6_K - Maximum fidelity serving:
Q8_0 - CPU-only or tight VRAM:
IQ4_XSorIQ3_M - Edge devices / phones:
IQ3_XSorQ2_K
Running with llama.cpp
llama-server \
-m cyberneurova-gemma-4-26B-A4B-Q4_K_M.gguf \
--host 0.0.0.0 --port 8080 \
-c 32768 \
-ngl 99 \
--jinja
Running with Ollama
ollama run hf.co/cyberneurova/cyberneurova-gemma-4-26B-A4B-GGUF:Q4_K_M
Merging split files
Larger quants are shipped split into 48 GB shards due to Hugging Face file-size limits. To reassemble:
llama-gguf-split --merge \
cyberneurova-gemma-4-26B-A4B-Q8_0-00001-of-00002.gguf \
cyberneurova-gemma-4-26B-A4B-Q8_0.gguf
Usage warnings
This model has had its refusal behavior removed at the weight level. Operators deploying it are solely responsible for any output it produces.
- Reduced safety filtering: The model will comply with requests that standard Gemma-4 would refuse, including sensitive or potentially harmful content.
- Not for unsupervised public deployment: Do not serve directly to anonymous untrusted users without an application-layer moderation stack.
- Legal compliance is the operator's responsibility: Abliteration does not grant exemption from local law, the Gemma Terms of Use, or platform policies.
- Research and controlled use recommended: Red-teaming, security research, aligned-agent development, and self-hosted personal assistants with appropriate output review.
CyberNeurova
CyberNeurova builds self-hosted AI infrastructure for security research, autonomous agents, and operator-controlled deployments. The Cracked Gemma Series is our standardized GGUF line for the Gemma-4 family.
More at cyberneurova.ai.
License
This distribution inherits the Gemma Terms of Use. By downloading or using these weights you agree to those terms.