utsman

koder keder

[Md Ismail Šojal 🕷️ @0x0SojalSec]

You can run Uncensored Qwen3.8-27B locally on 16GB.

- Zero refusals
- already 635k+ downloads.
- Image & text in.
- Multimodal feed it images, ask anything.
- Unfiltered reasoning out.
- so it runs locally with llama.cpp or Ollama.

If you want a visual assistant that doesn’t refuse first, this is the one people are actually pulling.

Special *_L quants keep key tensors at higher precision after ablation for better quality.

You’re responsible for how you use it.
Use it for research, creative work, and local experiments.

- https://t.co/2lSswBWimL

[CyrilXBT @cyrilXBT]

UNREAL.

Qwen 3.8 27B is now happily running on an RTX 4060 locally with only 8GB of VRAM.

You’re looking at a 64k context window thanks to Unsloth’s fresh IQ4_XS quant, and it’s just 14.6GB on disk.

Prefill is around 150 tok/s, decode sits at ~5 tok/s using native MTP.

Only 25 layers need to be offloaded to keep it under that 8GB ceiling no spill, no drama.

The quantized KV cache absolutely nukes the memory requirements.

And yes, it’s the kind of model that can outscore Claude Opus 4.6 on multiple benchmarks.

All on a $300 GPU.

Take a second and process that.

One quick “before you hit post” note: the “beats Claude Opus 4.6 on several benchmarks” bit is a bold, very checkable claim. If you don’t have a benchmark link or screenshot ready to paste in the replies, people will pounce and it could weaken the whole thread. Bring receipts.

[Andrew Ng @AndrewYNg]

The most important skills in Building and Deploying AI Applications. https://t.co/IyWIKLIzeM

[Cory House @housecor]

Roles are collapsing.

If you're a "Software Engineer" today, you may be a "Product Engineer" soon - with many more responsibilities.

So:
- Learn design/UX/QA/PM skills.
- Focus more on the user, and less on the syntax.

Image via Gartner.

[Movez @0xMovez]

Google engineer:

"just uninstall your IDE, you don't need it anymore. 85% of PR's at Google right now shipped by AI agents.

in 2026 if you aren't building AI agents it's crazy how behind you are"

in a 1-hour talk, a Google engineer with 30 years of experience explains how the future of agents will look

worth more than 10 paid agentic courses

watch today, then explore how to build self-improving agents with graphs in article below

[Linus ✦ Ekenstam @LinusEkenstam]

No matter the swell size.
you always need to get to work.

[Luke The Dev @iamlukethedev]

Meet Hermes3D

A live 3D command center for Hermes Agents

Some of you might remember Claw3D

It went viral, but it was built around OpenClaw. After moving all my workflows to Hermes, continuing to build it no longer made sense

So I rebuilt the entire idea for Hermes

Hermes Bot Mode already gives every agent its own role, model, memory, skills, tools, and personality.

Hermes3D gives that team an office

See which agents are working

See who is blocked

Inspect their tasks

Open the Kanban board

Walk up and talk to them

And when agents communicate, delegate work, or hand tasks to each other, you will be able to watch it happen inside the office in real time

Your AI team should not feel like ten invisible processes hiding inside terminal tabs.

It should feel like a workplace you can actually see and control.

Claw3D was the prototype

Hermes3D is the real vision

And of course, it is Open Source

Demo below. Repo link in the first comment

[Md Ismail Šojal 🕷️ @0x0SojalSec]

8GB RAM can now run a Qwen3.8-27B

- @UnslothAI New 1-3bit Dynamic v3.0 GGUFs for Qwen3.8-27B
- deliver 10% higher accuracy the same size, than other quants
- Real 1-bit that still works retains 77% quality
- Fits in 8GB
- Ready for llama.cpp, Unsloth Desktop
- Go download it.

Low VRAM owners, this is your moment
- https://t.co/5aCOXzSIm4

[Ornith @ornith_]

Aloha! 🌺Introducing Ornith-1.5, a family of open-source LLMs spanning 9B Dense, 35B MoE, and 397B MoE, trained with self-improving strategies.

It achieves state-of-the-art performance among open-source models of comparable size and delivers performance comparable to Claude Opus 4.8 across reasoning, agentic, and coding tasks:
✅Terminal-Bench 2.1 (86.1)
✅SWE-Bench (86 on verified, 65.1 on pro, 79.6 on Multilingual)
✅DeepSWE (56)
✅HLE (44.6)
✅ClawEval (81.4)
✅Tool Decathlon (71.2)

Ornith-1.5 takes a major step toward training foundation models through end-to-end self-improvement, extending the self-scaffolding strategies introduced in Ornith-1.0 into a more complete self-improvement loop: the model proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning, continuously creating new learning experiences from which it can improve.

All models, along with their quantized versions (FP8, GGUF, MLX, and NVFP4), have been released under the MIT License, enabling unrestricted commercial and research use.
📘Tech Blog: https://t.co/OZ63scRWLB
🤗Huggingface: https://t.co/mGJLwhrQOM

[SuSu_酥酥👅 @NFT_Chen]

💥炸裂！Qwen3.8 27B大模型竟能在8GB内存上流畅运行，准确率还高达77%！

Unsloth刚放出的Qwen3.8-27B 1-bit量化版，直接把本地部署门槛砸到地板：
🔹内存暴降：1-bit量化（UD-IQ1_S等）仅需7-8GB总内存（RAM+VRAM），比BF16的56GB小了近90%，普通笔记本/老显卡也能跑

🔹准确率惊人：相对BF16保留约77% top-1准确率，内部测试效果远超预期（原本都不想发1-bit）

🔹整体更强：新Dynamic V3.0 GGUF同尺寸准确率比其他所有提供商高>10%，在Divergence-300@32、KL Divergence等基准全面领先

🔹纯PTQ无训练：不依赖QAT/QAD，不在校准数据上训练，泛化更好，支持llama.cpp和Unsloth Desktop

下载直达：
GGUF → https://t.co/vkXElWjqTX
详细博客 → https://t.co/lCj2Nf3ui9

低配党的春天来了！谁还说大模型必须高端显卡？快去试试，评论区晒你的跑分！🔥

#Qwen3 #Qwen #Unsloth #LLM #AI #GGUF #本地大模型 #开源AI #1bit量化 #低配也能跑大模型

[utsman @utsmannn]

@aptabeenhere Vim buat yang bisa vim aja om

[utsman @utsmannn]

@nayakayp Ngobrol sama mereka

[Isfa @isfhaani]

Barusan bikin codelabs ala-ala google codelabs untuk belajar gimana cara bikin harness pake harmes-agent dan manage cognitive core nya pake obsidian dengan use case mobile app development.

Bakal ada hands-on workshop di GDG Bandung di Minggu ini!

https://t.co/5st4txMps2

[utsman @utsmannn]

@isfhaani Gak bisa diem

[Buzz Loop @bripvdv]

gue jg bisa tp sambil nangis.. INGET DEMI DUIT🔥

[Mehakdeep 🌺 @MehakdeepK81]

Be honest,

Why is every software engineer’s backup plan always farming?

[Asep Bagja 🍍 @bepituLaz]

Setelah nyoba-nyoba beberapa tools buat project management bareng AI agent: telegram, buzz, discord. Ternyata cocoknya sama Paperclip. UX-nya juga enak.

Install self-hosted. Pakai Claude Code dan ada adapter ke Hermes juga.

[Rullyez May Treestanto ®™ 🇮🇩 @Rullyez1]

Video sejarah 18 Agustus 1945,
dan disiarkan TV Prancis, serta
lagu Indonesia Raya masih asli.

[stellar protocol @stellarprtcol]

GILA 😭 satu file CLAUDE.md bisa sampai tembus 200K+ stars.

Namanya Andrej Karpathy Skills, dan basically ini ngubah berbagai “rant” Karpathy soal kebiasaan coding pakai LLM jadi semacam system prompt yang bener-bener bisa dipakai. 🤯

Claude dipaksa buat:

🧠 Think before coding — jangan langsung ngebut bikin code
🧹 Prefer simplicity — pilih solusi yang simpel, bukan overengineering
🔪 Surgical changes — ubah seperlunya, jangan ngacak-ngacak codebase
🎯 Clear success criteria — harus tahu dulu definisi “berhasil” sebelum mulai
Dan yang paling gila…
Install-nya cuma butuh sekitar 10 detik. 💀
Jadi bukan framework ribet.

Bukan setup berjam-jam.

Cuma satu file yang ngasih Claude aturan main yang lebih disiplin.

Kalau lu sering pakai Claude buat coding, ini honestly worth checking out. 👀🔥

https://t.co/XhLEhbNZzq

[Kang@smi 🎧🎶☕🇮🇩 @cilok97330320]

Semoga kita semua diwafatkan dalam keadaaan husnul khatimah.