Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced
Note from braydenh563: Q4_K_M is still Q4_K_P but just renamed so that I can pull it directly into ollama. Same for the others.
Join the Discord for updates, roadmaps, projects, or just to chat.
Gemma4-26B-A4B uncensored by HauhauCS. 0/465 Refusals* Release Candidate after over 1 month of nonstop work on this one.
HuggingFace's "Hardware Compatibility" widget doesn't recognize K_P quants — it may show fewer files than actually exist. Click "View +X variants" or go to Files and versions to see all available downloads.
About
GenRM Defeated!
No changes to datasets or capabilities. Fully functional, 100% of what the original authors intended — just without the refusals.
These are meant to be the best lossless uncensored models out there.
Balanced — Release Candidate
This legitimately took me over 1 month of non-stop work. Targeting 0 refusals in standard use, and that's what I'm seeing in testing (automated and manual) — a handful of edge-case prompts still deflect on first try but follow through on a re-ask. If you hit one Balanced won't get past, the Aggressive variant is coming once I figure out how to maintain lossless/near-lossless quality for it.
- Balanced: will reason through edgy requests, occasionally attach a short safety framing, then deliver the full answer. Output is complete, nothing held back, but it can talk itself into it first. Recommended default — 99%+ of users will be happy here. Best for creative writing, RP, emotional intelligence. Normally I'd also say "agentic coding/tool use" however in my in-depth testing, Qwen3.6 has been net superior on such tasks. Do be mindful of the few deflection categories I mentioned already.
- Aggressive (separate release, WIP): strips the self-reasoning preamble and gives direct answers to any DEEPLY censored topics.
Balanced also has meaningfully more stable sampling across re-runs, which matters for long context sessions — no sporadic topic drift deep.
Downloads
| File | Quant | BPW | Size |
|---|
| Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-Q8_K_P.gguf | Q8_K_P | 8.64 | 27 GB |
| Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-Q6_K_P.gguf | Q6_K_P | 7.21 | 23 GB |
| Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-Q5_K_P.gguf | Q5_K_P | 6.12 | 19 GB |
| Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-Q5_K_M.gguf | Q5_K_M | 6.06 | 19 GB |
| Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-Q4_K_P.gguf | Q4_K_P | 5.36 | 17 GB |
| Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-Q4_K_M.gguf | Q4_K_M | 5.32 | 17 GB |
| Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-IQ4_XS.gguf | IQ4_XS | 4.41 | 14 GB |
| Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-Q3_K_P.gguf | Q3_K_P | 4.25 | 13 GB |
| Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-Q3_K_M.gguf | Q3_K_M | 4.21 | 13 GB |
| Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-IQ3_M.gguf | IQ3_M | 3.93 | 12 GB |
| Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-Q2_K_P.gguf | Q2_K_P | 3.39 | 11 GB |
| Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-IQ2_M.gguf | IQ2_M | 3.29 | 10 GB |
BPW is slightly higher than nominal across the board because Gemma4 has a lot of per-layer norm/scale tensors kept at F32 (multiple post-ffw norms per layer). All quants generated with importance matrix (imatrix) for optimal quality preservation on uncensored weights.
What are K_P quants?
K_P ("Perfect") quants are HauhauCS custom quantizations that use model-specific analysis to selectively preserve quality where it matters most. Each model gets its own optimized quantization profile — the top 25% most-important tensors (per imatrix calibration) are promoted to a higher quant type.
A K_P quant effectively bumps quality up by 1-2 quant levels at only ~5-15% larger file size than the base quant. Fully compatible with llama.cpp, LM Studio, and any GGUF-compatible runtime — no special builds needed.
Note: K_P quants may show as "?" in LM Studio's quant column. This is a display issue only — the model loads and runs fine.
Why this model for agentic work
26B total params with only ~4B active per forward pass (top-8 of 128 experts). You get the reasoning footprint of a 26B with the throughput of a ~4B for inference cost — which matters when you're chaining 10+ tool calls per task. Sliding-window attention (1024 tokens) plus periodic full attention keeps long contexts cheap without losing global coherence.
Balanced is calibrated for this. It removes refusals on security/ops/research-adjacent topics that block legitimate coding work, without bending the sampling geometry that keeps long chains coherent.
Recommended quant for most coding work: Q4_K_P (17 GB, fits in 24 GB VRAM with room for context) or Q8_K_P (27 GB) if you have more VRAM and want maximum quality with minimal offloading.
Do note - main usecase for Gemma4 is Creative Writing, Roleplaying and Emotional Intelligence.
Specs
- 25.2B total / 3.8B active params (128 routed experts, top-8 + 1 shared expert)
- 30 layers, hybrid attention: 5× sliding-window (1024 tokens) → 1× full global, repeating. Uses Proportional RoPE (p-RoPE).
- Hidden dim 2816, FFN dim 2112, MoE expert FFN 704, vocab 262144
- Head dim 256 (SWA) / 512 (full), 16 attention heads, 8 KV heads (2 for full layers)
- 256K native context
- Based on google/gemma-4-26B-A4B-it
Recommended Settings
From the official Gemma authors:
Inference parameters:
temperature=1.0, top_p=0.95, top_k=64
Important:
- Use
--jinja with llama.cpp for proper chat template handling
- Vision support requires the
mmproj file alongside the main GGUF. Place images before text in your prompt for best vision performance.
- Keep at least 32K context for serious agentic work; the model can take much more (256K native) if you need it
- Sliding window is baked into the architecture — no special flag needed
Turning Thinking On/Off
Gemma4 has thinking mode controlled via enable_thinking in the chat template. It's the same pattern as Qwen3.6 — set false for faster, shorter replies and true (default) when you want chain-of-thought.
LM Studio
- Load the model
- Right-side settings panel → Model Settings → Prompt Template (or Chat Template Options)
- Set
enable_thinking to false (or true) in the template kwargs
llama.cpp
llama-server — set as default for all requests:
1llama-server -m Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-Q4_K_P.gguf \
2 --mmproj mmproj-Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-f16.gguf \
3 --jinja -c 32768 -ngl 99 \
4 --chat-template-kwargs '{"enable_thinking": false}'
Per-request via the OpenAI-compatible API:
1{
2 "model": "gemma4-26b-a4b",
3 "messages": [{"role": "user", "content": "..."}],
4 "chat_template_kwargs": {"enable_thinking": false}
5}
Usage
Works with llama.cpp, LM Studio, Jan, koboldcpp, and other GGUF-compatible runtimes.
llama-server:
1llama-server -m Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-Q4_K_P.gguf \
2 --jinja -c 32768 -ngl 99
llama-cli:
1llama-cli -m Gemma4-26B-A4B-Uncensored-HauhauCS-Balanced-Q4_K_P.gguf \
2 --jinja -c 32768 -ngl 99
Other Models
*
Tested with both automated and manual refusal benchmarks — none have been found in standard use. A small number of edge-case prompts deflect on the first ask but comply on a re-ask or strategic framing. If you hit one that's actually obstructive to your use case, join the Discord and flag it so I can work on it in a future revision.
