Breaking

SuperGemma 26B, an Uncensored Fine-Tune of Google's Gemma, Gains Traction Among Local AI Enthusiasts

July 13, 2026 at 06:45 EDT

Open Source
Foundation Models
AI Agents

A community-built variant of Google's Gemma called SuperGemma4-26B-Uncensored is drawing attention for running entirely on personal hardware with its safety filters stripped away. Released on Hugging Face by a developer identified as Jiunsong, the model is built on Google's Gemma-4-26B-A4B-it and marketed as "fully uncensored," designed to respond without the refusals typically baked into commercial assistants.

Early 2026 · Open-Weight Community

A 26B "Uncensored" Gemma That Runs On Your Own Machine

SuperGemma4-26B-Uncensored, a community fine-tune of Google's Gemma, strips out safety refusals and ships in local-first formats — GGUF, MLX 4-bit and a one-command Ollama build — putting offline, filter-free inference within reach of consumer hardware.

26B

Parameters — compact enough to run locally

256K

Token context window

~17GB

Q4_K_M build size on disk

~0%

Reported refusal rate in testing

Memory footprint fits consumer hardware

RAM/VRAM needed to run the recommended quantization — compared against a typical high-end consumer GPU.

16–22GB

SuperGemma to run

24GB

High-end GPU capacity

Also runs natively on Apple Silicon Macs via unified memory.

Three ways to run it locally

GGUF v2

for llama.cpp

→

MLX 4-bit

for Apple Silicon

→

Ollama

one-command install

What users report

Noticeably sharper than stock Gemma-4 26B
Substantially faster prompt processing
Strong on coding, tool use and agentic tasks
~3,550 downloads · 83 likes

The caveats

Thin documentation, no fine-tuning detail
No published metrics to verify claims
Still bound by Google's Gemma license
Removing guardrails reopens safety questions

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.