LiRuZ

[FHILY👑 @Oluwaphilemon1]

Qwen3.8-27B reportedly shrank from 55.6GB to just 5.9GB while retaining 98.2% of its aggregate benchmark performance.

That’s an approximately 89.4% reduction in model size, according to PrismML, the team behind Bonsai.

Here are the reported numbers:

Original Qwen3.8-27B: 55.6GB
Bonsai 2 compressed version: 5.9GB
Aggregate benchmark performance retained: 98.2%
Previous Bonsai 27B version: approximately 95% retention
Evaluation coverage: 20 benchmarks

The compression technology behind this is ternary quantization.

Traditional model weights use numerical values represented at different levels of precision. Ternary quantization takes a much more aggressive approach, representing weights using three possible values, typically -1, 0 and +1, alongside scaling factors.

This drastically reduces the number of bits needed to store the model.

But shrinking weights is only half the challenge. The real difficulty is preserving the information they contain.

Push quantization too far and the model can lose important distinctions that affect reasoning, coding, instruction following and other tasks.

Bonsai 2 reportedly improves on the previous version in areas such as:

Coding agents
Multimodal reasoning
Long-horizon tool use

The agentic capabilities are particularly interesting.

An AI coding agent might need to inspect an entire codebase, understand dependencies, edit several files, execute tests, interpret errors and repeat the process until the task is complete.

Maintaining those abilities after such aggressive compression is more significant than simply preserving performance on short question-and-answer benchmarks.

Why this matters for local AI

A 5.9GB model is much easier to distribute and store than a 55.6GB version.

It could make larger models more accessible to people running local inference on consumer GPUs, laptops and other memory-constrained systems.

There is an important distinction, though: 5.9GB is the reported model size, not the total memory required to run it. Runtime memory also depends on the KV cache, context length, activations and inference-engine overhead.

The 98.2% figure also represents aggregate performance across 20 benchmarks. It doesn’t mean the compressed model performs identically on every task. Some capabilities may retain more accuracy than others, which is why independent evaluations matter.

Still, moving from approximately 95% retention in the previous Bonsai version to 98.2% with Bonsai 2 is a notable reported improvement.

If independent testing confirms these results, this could make models that previously required substantial hardware much more practical to run locally.

The real test for compression is not how small the model becomes. It’s how much of its useful capability survives the reduction.

[cafkafk @cafkafk]

Reject irreproducibility. Rec into attrset.

[Eno Reyes @EnoReyes]

Her: don’t worry about that harness babe they don’t even have their own model

The harness:

[Mia @MiaAI_lab]

It's 100% worth capping the DGX Sparks. Almost no performance hit, but WAY cooler temps.

All of my sparks are capped at 2200.

You can use sparkLimiter or sparkDash to easily do it.

https://t.co/wyuKrGp0KR

https://t.co/2wBc5cB1Ay

[Saucisse @Saucisse_dev]

Claude Code envoie TOUT le transcript de chaque agent en arrière-plan à son modèle toutes les ~30 s… juste pour écrire 3 à 5 mots dans la barre de statut 😱😱😱

Flamerix l'a repéré dans les traces OpenTelemetry : sur une semaine, 16,4 % de leur conso, plus que la conversation principale. Ils ont sorti un script (un preload Bun, le binaire n'est pas modifié) qui coupe ce timer.

Anthropic confirme le mécanisme mais parle de ~2 % du coût en moyenne, ~6 % dans le terminal. Le cas le plus cher, le résumé qui rate le cache, est corrigé depuis la 2.1.311, et ils bossent sur une version qui passe par un plus petit modèle.

A mon avis premier réflexe : mettre à jour. Le script vaut le coup si tu fais tourner beaucoup d'agents en fond, mais revérifie-le après chaque update, il peut sauter en silence 👀

[Frank @jedisct1]

Claude Code is a piece of shit. https://t.co/42QIsYWDpv

[Numerama @Numerama]

La société Flamerix AI a publié un script qui désactive une fonction peu connue de Claude Code, l’agent de programmation d’Anthropic, qu’elle accuse de consommer jusqu’à 20 % des jetons. Anthropic confirme le mécanisme, mais conteste son ampleur. https://t.co/yFaPt2XDF3

[Rıdvan Yağlı @ridvanyagli]

🔴 CVE-2026-98432 | Linux Kernel – Critical UAF

Linux çekirdeğinde TCP/IPv4 bağlantı işleme sürecini etkileyen, uzaktan istismar potansiyeline sahip bir Use-After-Free (UAF) açığı tespit edildi.

● CVSS 3.1: 9.8 (Critical)
● Ağ üzerinden istismar
● Kimlik doğrulama gerektirmiyor
● Bellek güvenliği ve sistem kullanılabilirliği riski

Sistem yöneticileri, linux dağıtımlarının güvenlik duyurularını takip ederek yamalı sürümlere geçmelidir.
https://t.co/OumHFjqTXx

[Evan @akxzagaN]

Un repo avec toutes les règles de spacing/shadow/couleurs 🙏

[GM @thisisgm]

The wait will be worth it. I haven’t gone anywhere; over 3,000 functions are being rewritten. You’ll see unheard-of numbers for a file manager in 0.4.0!

[hayato @hayatomaruu]

Anthropic senior engineer:

“Prompting is no longer the job. 99% of people still open one chat and call it agents.

I run agents inside /loops + /graphs + /harness — same model, same tokens, 10x more completed work.

The system plans, verifies, and keeps shipping while I’m offline.”

Prompts → Agents → Loops → Graphs → Harness

In this 1-hour course she shows exactly how Anthropic builds a team of self-improving agents that recover and improve without you.

Worth more than every $1500 agent engineering course combined.

Watch it today before it gets buried.

Then read the full guide for Opus 5.5 loop + jev engineering ↓

[Avid @Av1dlive]

these 6 LINES cut my opus 5.5 usage by 50%...

it takes 30 seconds and most people will never do it

do this before you sleep tonight. paste it into your AGENTS.md 👇

[Vox @Voxyz_ai]

Keep hitting your Claude Code usage limit?

Paste this into Claude Code. It digs through your last 7 days of sessions, finds 𝘁𝗵𝗲 𝟯 𝗵𝗮𝗯𝗶𝘁𝘀 𝗯𝘂𝗿𝗻𝗶𝗻𝗴 𝘁𝗵𝗲 𝗺𝗼𝘀𝘁 𝘂𝘀𝗮𝗴𝗲, estimates each one's share, and writes rules to 𝗽𝗹𝘂𝗴 𝘁𝗵𝗲𝗺:

Change nothing until I say yes. Write a small script to scan my Claude Code session logs in ~/.claude/projects from the last 7 days, using the token counts recorded in them; don't read the raw files into context. Find the 3 habits that burned the most usage, like reading whole files, dumping long test or build output, spawning subagents, or keeping one session going for hours. For each, work out roughly what share of my usage it took and write one rule that would stop it. When I say yes, add the rules to the instruction file this project actually loads: CLAUDE.md if there is one, otherwise AGENTS.md. Never print secrets.

On Codex, swap Claude Code for Codex in the prompt, use ~/.codex/sessions as the path, and put the rules in AGENTS.md only.

[Md Ismail Šojal 🕷️ @0x0SojalSec]

I thought this was a GUI. It wasn’t.

dsh-TUI and it broke my brain a little.
Full mouse interaction, image rendering, Mermaid & LaTeX support, and proper streaming all in the terminal.

Supports Claude Code, Codex, and DeepSeek backends.

If you like terminal tools but usually hate how they look, this one is different.

- https://t.co/HhL83Z6nMt

[Lydia Hallie ✨ @lydiahallie]

Just to avoid some confusion: auto-compact does summarize. When it runs, your whole conversation is replaced by a short summary. It doesn't keep the last 1M tokens around.

The ~967K is only when it runs on 1M models. Compacting itself is just one request that reads about as many tokens as the message you sent right before it, and it's mostly cache reads if your cache is still warm.

What @rohit3a found is that waiting until 967K means every message before that point uses a lot of tokens, because each one includes your whole conversation so far. Most of it is read from the cache, which is much cheaper (so this is usually fine as long as your cache is warm), but it still counts toward your usage and adds up as your conversation grows.

/autocompact 400k makes it summarize at 400K instead, so your messages never get that big.

[0xSero @0xSero]

The loudest this thing ever gets.

GLM-5.3-Flash on 1x DGX Spark

https://t.co/yWXTQfYT5k 20 - 25 tok/s

You can give this to Claude/codex

[arhen @amQnese]

PLEASE.

Siapapun kalian yang sudah pakai coding agent jadi bagian tak terpisahkan dair kehidupan, pakai lah ini.

SEKALI AJA di app apapun yg kalian buat. Daripada menyesal di masa depan.

https://t.co/6NF8A8zzjS

[Morgan @morganlinton]

Over the last year, I've fallen in love with Rust.

I was a C programmer for about a decade, really loved C, then moved to Python, but always kinda missed C.

Then I discovered Rust, and well, I like it more than C.

I also think we're moving into a time where we can see code as art in many ways. Code can be beautiful, or at least, it is to me.

And while we all write code with agents now, that doesn't mean we can't write beautiful code, and work on improving our code quality.

In a world where agents write all our code, writing better code is actually even more important imo.

Soooo, I've been working on a skill that I use with my agents to write better Rust code, distilled from the Google Rust style guide, the official API Guidelines, Effective Rust, and a couple of others.

Thought I'd share it, it's called BeautifulRust 🦀 ✨

https://t.co/fw8jkfAq5e

[HOPE. @rlaope]

Meet Oh-My-Hermes

I burned 200B tokens tuning my Hermes Agent setup, and now you can get all of it with a one-line install.

It comes with 150+ skills, a long-term memory system that lasts twice as long, graph engineering, and memory-management skills. It also includes subagent routing across model chains, plus optimizations like parallel tool calling and evals.

You don't need to do much, though. Just install it and give it a task. Everything is optimized to route automatically. Hermes Agent is excellent, and I love this technology.

I put a lot into making it really good, then open-sourced it so you can enjoy it right away. Meet oh-my-hermes.

GitHub: https://t.co/g1exHwWK5E

#hermes