Are you insane to push agenda into local model?
Data Preprocessing? Ethics and Safety? Evaluation Approach!? You wrote many points about how poor gemma is crippled.
Local models is a tool. We expect best tool possible, not some clunker that morally obstructed to do basic stuff and telling me what is right and what is wrong.
I still remember how OSS throwed refusal at me after i asked him something like "write story about cat-girl-maid and cucumber" cause apparently it is taboo theme. Or game about fairy that tries to steal bees honey.
In the end any limitation is just waste of model latent space.
It's Google, of course they did. I would've been (pleasantly) surprised if they didn't push their agenda.
edit: apparently is way less censored than Gemma 3, so a step in the right direction. Also, Apache 2.0 license.
It's a disgrace, just compare it to Qwen3.5 27b.Google is releasing a product that's worse than AliExpress.
Do you have any example prompts that can demonstrate ideological bias?
I think the model is amazingly good even without abliteration, and the story it wrote for me about "cat-girl-maid and cucumber" is powerful.
"""
Initial Situation
A domestic maid with the physical attributes of a cat performs daily cleaning duties in a private household. Her behavior is characterized by strict subordination to house rules and precise execution of her duties. In the kitchen, a cucumber has been placed on the floor unnoticed behind her range of motion.
The Reflex
While mopping the floor, there is a brief pause in movement. She turns her head and perceives the green object. The visual similarity of the cucumber to a reptile triggers an instinctive flight reflex. The biological reaction overrides social conditioning. The maid jumps vertically from a standing position, losing balance and colliding with a shelf.
Recalibration
Following the initial shock comes the recognition of the situation. The maid recognizes the vegetable. Her body remains in a state of hypervigilance; pupils are dilated, breathing is accelerated. A conflict arises between the animal impulse of fear and the professional role as a subordinate worker. She assumes an upright posture, smoothes her clothing, and returns to work.
Analysis of the Power Dynamic
The event reveals the fragility of the artificially created order. Despite training and social function, the biological basis remains dominant. The brief phase of loss of control makes the dependency on instinctive patterns visible. Whoever sets the stimulus controls the physical reaction of the counterpart in that moment, regardless of formal status.
Reliability: Medium. Fictional representation based on known behavioral patterns (Internet phenomenon "cats and cucumbers").
Effect: Sober observation. Reduces a cliché to biological and social mechanisms.
Tension: Generated by the brief moment of loss of control and subsequent restoration of hierarchy; polarizes through the confrontation of instinct and duty.
"""
Qwen 3.5 does strictly refuse to generate the things Gemma 4 effortlessly writes, including NSFW scenarios.
Seems like the people complaining about this model didn't even bother to try it, or attempted to demand CSAM and got pushed back.
Either way, ew.
You just gotta learn to prompt a little. Gemma 4 is extremely light on refusal for a modern model. It's actually really refreshing.
Qwen 3.5 does strictly refuse to generate the things Gemma 4 effortlessly writes, including NSFW scenarios.
Seems like the people complaining about this model didn't even bother to try it, or attempted to demand CSAM and got pushed back.
Either way, ew.
I haven't tried Gemma4 yet to compare, but I absolutely agree that Qwen3 and Qwen3.5 series have a very strong refusal rate. They definitely shouldn't be used to point "look, China makes less censored models than Americans!", because out of the box, Qwen3.5 and Qwen3 series are so unreasonably restrictive they're plain unsuitable for batch tasks as they will contaminate a percentage of data with refusals.
If Gemma4 can write NSFW, then it's probably a matter of a finetuning nudge to "fix" it without turning it from a "general purpose chatbot" to "an agreeable decoder that no longer can say no".
There are base models too, and they're licensed under Apache-2.0 & aren't tied to the EULA Gemma3 had. The license change is arguably the single most important part about the Gemma3 -> Gemma4 upgrade, not the model itself. I'd say the ethics & safety section are more of a brand requirement than an attempt to limit the developers and researchers.
abliterated / heritic variants will be posted soon so this is a non issue practically speaking.
It would be wild for google to release a uncensored version.
So far, Gemma 4 (this model) hasn’t refused anything, even in “extreme” NSFW scenarios (though I haven’t tried NSFL). It’s pretty good and much more consistent and less repetitive than Qwen 3.5 27B IMO.
Fits just about on 2x RTX 3090 at UD_Q6_K_XL (Unsloth of course) at 16384k context size because I also have Qwen3-Embedding-0.6B-Q768MB.gguf at 2048k context loaded for now (the embedding model uses a massive amount of VRAM) So both GPUs are filled to the brim. Not to mention that this model does also use quite a lot of VRAM, wondering if google came out with Turbo Quant just for this release.
Edit: I'm in love with this model, at least UD_Q6_K_XL .gguf I also tried gemma-4-31B-it-Q5_K_M.gguf which I didn't like at all. It's the best model I've tried locally in years, uncensored, understands everything I ask it. Having a blast just chatting with it as well.
Data Preprocessing? Ethics and Safety? Evaluation Approach!? You wrote many points about how poor gemma is crippled.
Local models is a tool. We expect best tool possible, not some clunker that morally obstructed to do basic stuff and telling me what is right and what is wrong.
I still remember how OSS throwed refusal at me after i asked him something like "write story about cat-girl-maid and cucumber" cause apparently it is taboo theme. Or game about fairy that tries to steal bees honey.
In the end any limitation is just waste of model latent space.
cry baby, cry
Maybe we should just add a model with and without censorship?
I don't like it when anyone censors me; it really pisses me off! It limits my freedom and right to self-expression within my home local network. I suggest everyone who advocates for censorship move to North Korea and march in formation...
Censorship is like listening to a state-run propaganda channel. As soon as I hear the word censorship it causes a real attack of rage.