Observer

Blockchain & AI degen.
Building the future.
Ex-@arkham

[Observer @0bserver]

My Qwen3.8-27B quants just hit 1.2M downloads in 2 weeks.

The mission stays the same: remove refusals—not the model.

Now for Qwen3.8 Flash Next:
Minimum guardrails. Maximum consistency.
2% strict refusal rate.

RVN is live:
https://t.co/LLRppg0Bhv

[Observer @0bserver]

Abliterated with heretic ARA: refusals 95 → 0 on 100 test prompts, normal answers unchanged (97/100), KL 0.0115.
Method and all numbers are in the model card.

[Observer @0bserver]

Our open models just passed 4M downloads on Hugging Face. Here's the next one.

Nanbeige4.3-33B, uncensored, on a single 16 GB GPU.

Instead of one quant type for the whole model, we picked the best one for each layer.

11.6 GB GGUF, only +3.8% perplexity vs. full precision. A plain IQ2_S of similar size loses 14%.

GGUFs for 16, 20 and 24 GB cards 👇

[Observer @0bserver]

Which file for your GPU:

16 GB → RCO 2.60 (+3.8% PPL)
20 GB → IQ3_S (+0.6%)
24 GB → IQ4_XS (+1.8%)

Plain IQ2/IQ3 quants are in the repo too, if you want to compare.

[Observer @0bserver]

@sudoingX My model's outputs are much better with multilingual use btw. Wanted to give a heads up. And thx for the share :)

[Observer @0bserver]

I need some motivation to back up all my Linux machines and install Omarchy. I know I will do it someday, but idk when lol.

[Observer @0bserver]

@Heidagger95 Thx man, but also thanks to you :)

[Observer @0bserver]

The latest RVN Uncensored Qwen 27B has over 2.3 M all-time downloads.

Someone in the comments asked for GSQ-RCO, so I made the same weights to fit better on a 16GB card.

https://t.co/3WrwooWpwC

IQ3_S: 11.80 GB, +0.95% PPL vs F16.

Uniform IQ3_S: 12.42 GB, +2.9%.

851 tensors, each with its own type. Start at IQ3_S on 16GB.

[Observer @0bserver]

@FvSvY

[Observer @0bserver]

My Qwen3.8-27B quants just hit 1.2M downloads in 2 weeks.

The mission stays the same: remove refusals—not the model.

Now for Qwen3.8 Flash Next:
Minimum guardrails. Maximum consistency.
2% strict refusal rate.

RVN is live:
https://t.co/LLRppg0Bhv

[Observer @0bserver]

@ADLXBT @typesafeai the 2% is the epic fails with confidence lol when it misses, it doesn't hedge, it comes back at 1.00 on the wrong label. Im ok with an uncertainty to a point, but if I see a confidence above 0.98, I expect it to be a hit, not a miss.

[Observer @0bserver]

Spent the night benchmarking Jev from @typesafeai.

Ups: 20 extra questions cost +47ms; 500+ api calls, barely spent 0.03 $

It caught a fake-approval injection on my own fine-tuned agent kernel. I'd say %98 success rate (spent so much time with jev lol)

Downs: it reported confidence 1.00 on a wrong answer, so no, you can't gate on it. YET.

However, I'm impressed for a v1!

Congrats @CompleteSkeptic and the team

[Observer @0bserver]

@sama @OpenAI Ok ok, we are good now. A few banked resets would be nice tho lol

[Observer @0bserver]

Hey @sama @OpenAI, my account got banned for "Prohibited Biological Use."

Never made a bio or weapons query. Literally checked all my logs; zero.

Actual traffic: a legal assistant processing Turkish court decisions, which run across "weapon" constantly in a legal context. That's what your classifier flagged.

[Observer @0bserver]

@thsottiaux So you also decided to trim heavy users ha?

https://t.co/dCZr6MGMdz

[Observer @0bserver]

Hey @sama @OpenAI, my account got banned for "Prohibited Biological Use."

Never made a bio or weapons query. Literally checked all my logs; zero.

Actual traffic: a legal assistant processing Turkish court decisions, which run across "weapon" constantly in a legal context. That's what your classifier flagged.

[Observer @0bserver]

@0808 @sama @OpenAI Duhh… that looks like a regular prompt to me. Idk what their problem is lol

[Observer @0bserver]

@QiuTing395298 @sama @OpenAI I might have asked how to adjust the dosage of paracetamol according to my kids weight a few times lol

[Observer @0bserver]

@mktpavlenko @sama @OpenAI The worst part is, they are failing it. lol

[Observer @0bserver]

Yeah... That's why we use local models 😂

[Observer @0bserver]

@alexandr_wang Any updates for glimmer?