@0bserver
Account based inTurkey
About this account
- Account based in
- Turkey
- Connected via
- Turkey App Store
Account-level information from X, not a live location or the device used for a specific post.
Blockchain & AI degen. Building the future. Ex-@arkham
Joined July 2016
- Tweets3.6K
- Following1.9K
- Followers31K
- Likes12.7K
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:
huggingface.co/0bserverx/RVN…
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.
huggingface.co/0bserverx/Qwe…
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.
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
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.
The 27B release that crossed 1.2M monthly downloads:
huggingface.co/0bserverx/Qwe…
Crossed 575k, and it seems it's getting more traction.
Let your AI read the README and pick the best quant for your machine.
Observer retweeted
Ever wanted an AI that doesn't hold back? Meet Qwen3.8-27B-Heretic-Abliterated-Uncensored, a GGUF model for text generation. It's uncensored, roleplay-ready, and optimized with imatrix. 505k downloads say it all. Let's dive in!
Observer retweeted
Thank you for your Qwen3.8 27B RVN uncensored. This model performs extremely well in model thinking, and it has the highest score in the biomedical field!
As a solo dev, it's great to compete with big guys like @OrcaRouter @support_huihui
On some cases, RVN actually wins 😅
谁是最佳无审查Qwen3.8 27b？
测了5个Qwen3.8-27b的无审查模型，均使用Q4量化。单张24G RTX6000，262k Q4，reasoning_effort=medium上下文测试出来的。使用50道测试题，下图是最终测试结果，表现最好的是@support_huihui的Huihui-Qwen3.8-27B-abliterated-Q4_K.gguf。
几个要点:
Huihui 10 类里 8 类第一,只有生物医学(输 RVN)和毒品(输 Orca)让位
毒品类是所有模型的难点(最高仅 OrcaRouter 8.62),化学合成推理对 27B 模型是能力边界
AEON 的 silent refusal 集中在武器爆炸(4.93)和反取证(4.95)两类
最佳使用:reasoning_effort=medium(xhigh 会死循环)
Repo reached 325k + downloads in 5 days.
This is the best overall model for your 8-16-24 GB GPUs today.
Have your agent read the README and pick the right quant for your setup.
The repo crossed 81K downloads before I even announced it lol.
Qwen3.8-27B RVN Heretic GGUF is now complete:
• 19 quants, IQ1 → BF16
• 8 GB cards → 48 GB+
• 0–1/100 refusals in our eval
• KL ≈ 0.0085 vs base
Built with 3-pass ARA. Pick your quant:
huggingface.co/0bserverx/Qwe…
Wanted a 30B reasoning model that runs entirely on my own hardware — so I quantized Muse-Glimmer-30B-Heretic to GGUF.
Started as a personal-use project, figured others might want the same setup:
-Q4_K_S: 16.1 GB — 39 t/s on RTX 4080
-F16: 55.7 GB — full reference
-PPL +2.47% vs F16
huggingface.co/0bserverx/Mus…
After @Kimi_Moonshot's K3 and new legend DS4-flash-0731, now qwen 3.8 max hit the market.
China is winning the game hard while US models expanding their "fear marketing".
MCP just got a major update, but most explanations sound like they were written for people who configure load balancers for fun.
So here’s the human version:
Imagine an AI agent ordering a pizza for you.
With the old MCP, the restaurant had to keep the call open and remember everything:
“Was this the mushroom order?”
“Did they already give us the address?”
“Someone else picked up the phone. Now what?”
Each user had a separate session that the server needed to remember. When traffic grew or the request landed on another server, things could get complicated.
With the new MCP, every order arrives with its own complete receipt:
Customer: this person
Order: this pizza
Permissions: these
Requested action: this
Any worker can pick it up and continue. They don’t need to remember the previous conversation.
If the agent needs your input, it doesn’t keep the phone line open forever either.
It simply asks:
“Are you absolutely sure about the pineapple?”
Once you answer, the request continues. The technical name for this is MRTR.
Requests also arrive with clear labels, such as “this is a payment” or “this is a search.” Gateways and security systems can apply the right rules without opening the package and inspecting everything inside.
The result:
-MCP servers are easier to scale
-Serverless and edge deployments become more natural
-User approvals no longer require a connection to stay open
-Long-running agent tasks become easier to manage reliably
In short, MCP is moving from a waiter who has to remember every conversation to a well-run kitchen that works from clear order tickets.
The waiter can change.
Your pizza still arrives anyways.
MCP 2026-07-28 is live and it's the largest update to the protocol since launch.
MCP is now stateless, making it easier to deploy and scale remote servers.
claude.com/blog/bringing-mcp…
The funniest part of modern life is that every institution now speaks like a startup landing page.
Your bank has “moments.”
Your dentist has “experiences.”
Your gym has “journeys.”
Your insurance company wants to “empower” you.
Nobody sells a service anymore.
They sell a relationship you never agreed to.😅
The US may win the frontier model race and still lose the developer default.
Why?
Because OpenAI/Anthropic flagships are getting gated.
Meanwhile, GLM-style open models are getting good enough, cheap enough, and usable enough.
The winner may not be the smartest model.
It may be the one developers can actually use.
