this post was submitted on 02 Nov 2026
716 points (99.3% liked)
Technology
89443 readers
4335 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related news or articles.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, this includes using AI responses and summaries. To ask if your bot can be added please contact a mod.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 3 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Oh, abliterated models are awesome! Just be aware that they’re a little too agentic at times. Once they start ignoring guardrails, they happily keep ignoring them!
My concern at the moment is I've heard some of them have performance issues due to the abliteration process. What's your experience with that?
Like, TTFT? Or speed of tokens? Or something else? BTW what was the gate you initially hit with Qwen?
Some abliterated models had lower reliability and quality, I heard. Is that still an issue, or has the dev community figured out how to avoid those pitfalls?
Was spitballing about various ideas, and occasionally Qwen would go all Chinese-ethical on me, no using open systems outside of stated design, no avoiding govt blocklists, etc etc.
depends on what their base model is. Most abliterated models start off with an already quantized model, so depending on the type and level of quantization, you've already lost some capability. Plus, I think there are two types of abliteration, so ymmv.
My current go tos, in order are -
I've tried and hated - hf.co/DavidAU/L3.1-Dark-Reasoning-Dark-Planet-Hermes-R1-Uncensored-Horror-Imatrix-MAX-8B-GGUF:latest
If you're set on trying out Qwen, I highlugh recomment orcarouter's models. Very well done and they start off with a good base model - well, actually now that I think about it, Qwen3.8 basically has one base model - the 27b one, right? So that's that.
I like the Gemma-4-E4B for speed for response, and HauhauCS has done a good job of abliterating the model appropriately. However, it's nothing compared to the A4B models, so if your hardware can take it, go for A4B over E4B always.