Executive Summary
The past 24 hours saw major advancements in local AI and model compression, heavily focused on running large models (like Qwen 3.8 27B) on older, consumer-grade hardware such as the RTX 3060. Developer tools and agentic ecosystems are maturing rapidly, with significant updates to Claude Code, OpenClaw, and AmpCode enabling more autonomous workflows and multi-agent orchestration. Additionally, the AI community engaged in critical debates over quantization sweet spots (1-bit vs. 4-bit) and the security implications of major AI labs hiding Chain-of-Thought (CoT) reasoning traces.
Key Events
- PrismML's "Bonsai 2" achieves massive local performance: A ternary compression of Qwen 3.8 27B (5.9GB) successfully ran agentic loops on a 5-year-old RTX 3060 12GB card at 26+ tok/s, completing a full UI build in one shot. → link
- Claude Code gets major infrastructure updates: The tool rolled out support for
AGENTS.mdand a proper extension system, allowing developers to build custom workflows and better manage context. → link - OpenClaw introduces multiplayer mode and web app testing: Users can now ask agents to run web apps in "Crabbox" and display them via VNC or a portal, alongside new conversation sharing and team server features. → link
- Debate over quantization sweet spots: Engineers argued that 1-bit/ternary models like Bonsai are currently less efficient end-to-end than 4-bit quantized models, which align closer to the ~4.7 bits of information processed by human synapses. → link
- Security concerns raised over hidden CoT traces: Researchers warned that OpenAI and Anthropic removing reasoning traces from their models harms white-hat security auditing by obfuscating how the AI thinks. → link
Analysis
The overwhelming trend of the last 24 hours is the democratization of inference. Developers are pushing extreme quantization techniques to make 27B+ models run on standard consumer GPUs, challenging the narrative that massive VRAM is required for local intelligence. Simultaneously, the software development lifecycle is being heavily augmented by agentic tools, shifting the developer's role from writing code to orchestrating agents, reviewing PRs, and managing context compaction. Next, watch for direct benchmark comparisons between 1-bit/ternary models and their 4-bit quantized counterparts, as well as the rapid proliferation of plugin ecosystems for agentic coding tools like Claude Code and Jev.
Tweet Feed
Local AI & Model Compression
@sudoingX · 2026-09-19T18:57
in 2026 qwen 3.8 27b runs an agent loop on a five year old rtx 3060. it says where local ai is headed. we need more of this energy and innovations... → link
@sudoingX · 2026-09-19T18:00
dear rtx 3060 owners, and every 12gb card behind it. bonsai 2 27b dense went from 26 to 40 tok/s tonight on the same 5.9gb file, and doubled to 50 tok/s with the mtp head grafted back, not one output changed. → link
@sudoingX · 2026-09-19T17:15
your rtx 3060 was running bonsai 2 at 26 tok/s this morning and now does 40 tok/s, i spent the day inside the prismml llama.cpp fork so your rtx 3060 can rejoice. the kernel that reads the weights was leaving two thirds of the card's threads idle... anyone can run it tonight... → link
@sudoingX · 2026-09-19T11:00
steam's hardware survey for august... most of you are sleeping on the card in your own setup. the number one gpu on that list ran a 27b model as an agent last night, one prompt, 77 minutes, a working gpu monitor out the other end... → link
@sudoingX · 2026-09-19T09:31
you did not need the 4090. a $200 used 3060 runs this 27b at 26 tok/s and a 4090 runs it at 91 tok/s... the 24gb tier is 5% of steam. the other 95% got told for two years they needed it. your card was never the problem, the file was. → link
@sudoingX · 2026-09-19T08:59
the timeline spent two days saying bonsai 2 cannot build. here is 1 hour 24 minutes of it building, one shot, from one paragraph, on an rtx 3060 12gb... what you are watching is a 5.9gb ternary compression of qwen 3.8 27b... → link
@sudoingX · 2026-09-19T00:02
bonsai 2 27b just built this from one paragraph of prompt in one shot, all of it out of a 5.9gb file on an rtx 3060 12gb. i did not expect frontend taste at this size. small models usually get the logic right and the layout wrong... → link
@sudoingX · 2026-09-18T19:42
bonsai 2 27b on an rtx 3060 12gb, the full receipt sheet. save this one... speed by depth, then what context costs, live server, thinking on, real sessions → link
@alexocheema · 2026-09-18T22:46
As much as I'd like this to work, 1-bit / ternary models like Bonsai are (today) a dud. The sweet spot appears to be ~4 to 5 bits per parameter. That's what @UnslothAI find to be the "knee" of the pareto curve... → link
@gospaceport · 2026-09-19T00:22
Bonsai 2 27B Ternary has mega hype, but does it match Qwen 27B FP16 to 98%? Is it close to 98%? I ran the same arcade box prompt on both in Hermes Agent. Let me know what you think below! /🧵 → link
Developer Tools & Agents
@iamdevloper · 2026-09-19T07:39
Jev making its blazingly fast classification responses → link
@kunchenguid · 2026-09-19T19:36
almost every day i hear people ask "when should i /compact my session"... but we have Jev now! introducing compact-adviser - an agent plugin you can use in claude and pi today to help determine whether you're likely at a task boundary that's safe to compact → link
@RayFernando1337 · 2026-09-18T19:05
RT @trq212: We're adding support for AGENTS.md to Claude Code. Starting today in version 2.1.277, if there is no CLAUDE.md in a folder, C… → link
@badlogicgames · 2026-09-18T19:10
and all it took was to add a proper extension system to claude code (sorry thariq, could not resist :D) → link
@steipete · 2026-09-19T01:03
Just landed a new feature, in the latest OC you can now ask your agent “Run this [web app] in crabbox and show me [vnc / a portal]” This was already working for cloud sessions... → link
@steipete · 2026-09-18T22:19
RT @openclaw: The Death of the Meat Proxy is here. Welcome to Multiplayer Mode from OpenClaw... → link
@sqs · 2026-09-19T09:59
New Amp CLI --attach: Attach files to a new orb thread:
amp --attach ./report.pdf -ox "Extract data from this PDF"→ link
@FinansowyUmysl · 2026-09-19T09:01
Badanie mówi, że problem mają uczelnie z dostosowaniem programu do zmian, juniorzy ze znalezieniem pracy i programiści z review kodu. A moim zdaniem, w ciągu roku rozwiążemy problemy z review kodu w branży - powstaną przystosowane do tego narzędzia (patrz Jev). → link
@badlogicgames · 2026-09-18T19:08
recommended reading, with the caveat that i think the article goes too hard on bend 2. the kernel: you often only understand a problem by experiencing the journey, i.e. building a solution for it. if you hand that journey off to an agent, you will get something. but not necessarily something good... → link
@LinusEkenstam · 2026-09-19T16:16
Incumbent design software is being replaced by vibe-coded jigs: tiny, purpose-built tools with a TAM of one. But it also gets replaces by new software, that runs in the browser, that challenges the status quo. → link
Hardware, Inference & API
@thdxr · 2026-09-19T14:38
also we we've been working for months to reproduce deepseek quality of inference. it is difficult. like $100M budget + right connections + getting the right 10 people in the world to help you difficult → link
@thdxr · 2026-09-19T14:34
btw another form of scam, i see inference providers claiming "99% cache rates" the only provider in the world that hits that right now is deepseek. so this means they are wrapping deepseek... → link
@TheAhmadOsman · 2026-09-19T06:00
Kernels implement specific tensor programs, not models. Model semantics → shapes → kernels → instructions → runtime → silicon. Your GPU behaves like a very expensive space heater without these things handled well → link
@TheAhmadOsman · 2026-09-19T02:02
Based on my experience with DeepSeek V4.1 Flash, GLM 5.3, and Kimi K3. - DeepSeek is excellency itself when it comes to pre-training. - Zhipu and Moonshot are better at post-training though. We're so lucky to have such amazing opensource players btw → link
@ivanfioravanti · 2026-09-19T10:29
Qwen-Image-2.1 experiments on Apple Silicon (M5 Max, MLX, bf16) with my custom MFLUX. 1600×672 (2.39:1), fixed seed: 10 steps → ~26s 20 steps → ~44s... → link
@ivanfioravanti · 2026-09-19T01:12
LLM Context Benchmark Splash inference engine by @inco_ai running on M5 Max 128GB macOS 27 in high power mode... → link
@ivanfioravanti · 2026-09-18T22:24
On the same theme: a new vertical inference engine for Apple Silicon will be released really soon (not by me) and it's really fast! Be ready! 🚀 → link
@thdxr · 2026-09-19T02:01
compute situation is crazy. i have an opencode session doing work with trainium and it had to make a reservation for capacity on saturday. so it scheduled a task to wake itself up then so it can continue working → link
@victormustar · 2026-09-19T17:42
RT @trycua: 1/ Introducing CUA-S1: a family of System One Models, small, specialized, and built for computer use. Today we're open-sourcin… → link
@victormustar · 2026-09-19T08:16
deepseek v4.1 flash is so good as a daily driver… I wish more people were aware. I feel it’s mostly a packaging thing for it to get huge adoption → link
Security, Ethics & Research
@TheAhmadOsman · 2026-09-18T23:05
OpenAI & Anthropic removing CoT/reasoning traces is actually a bad thing for security. How can White Hats identify security concerns within the models if they cannot see how the model thinks? If these companies are serious about security they wouldn't do security via obfuscation → link
@TrungTPhan · 2026-09-19T17:21
DraftKings spends $1B+ a year on sales & marketing... In 2023, data science team built ML model to find gamblers likely to lose money. It also created a “risk score” to flag gamblers “headed for trouble” (but DraftKings “sidelined” the tool): → link
@victormustar · 2026-09-19T10:01
Interested in AI and lunar science? The NASA-IBM Lunar Foundation Model, supported by USRA’s planetary science expertise, is now available on Hugging Face 🌘 → link
@MilksandMatcha · 2026-09-18T20:14
The global community benefits when labs share strong open models. @seanlie also sees a strategic challenge in how much of that progress is coming from Chinese labs. He argues that staying competitive in models and the hardware behind them will require a national effort... → link
@MilksandMatcha · 2026-09-18T19:22
They just launched a completely FREE AI browser agent to protect sensitive data that you send to agents like ChatGPT Claude or Gemini. The extension anonymizes any personal information in your initial prompt... → link