Minimizing structural material count and mobile ion displacement optimizes transport kinetics and mechanical durability, bounded only by fundamental charge cap.
Worms, Germany
Joined March 2014
- Tweets2.9K
- Following191
- Followers189
- Likes1.8K
RTX 3060, Ryzen 3800X, 2021 hardware. It runs a 128K context coding agent that reads git diffs and spots a -dirty version tag before I deploy it. Cloud not required.
GPU: @nvidia
CPU: @AMDRyzen 7 3800X
Model/quant: @Alibaba_Qwen 3.8-flash-next-iq3_xxs
Engine: Strata by @coldniko
GPU tuning: github.com/Zagitta/gpu-unloc… – at Worms, Deutschland
This is #strata ! - My trusty RTX 5060 12GB no serves Qwen 3.8-flash on my dev homestation with nearly 60toc/s and prefill at 880k is usable.
Checkout engine v0.1.52 features. Next is coding benchmark on michael.wegener.engineering/… – at Worms, Deutschland
Michael Wegener retweeted
First 4×V100 + Strata + DeepSeek benchmark is wild: even at 1M context, prefill 1k TPS beats llama.cpp + Qwen3.8-Flash by literally 4×, with no tuning yet. That’s why I waited for Strata—several engines can run V100 + DeepSeek V4.1 Flash, but honestly, I trust @coldniko (even if he forgot to follow me back 😁).
Qwen4 27b incoming. – at Worms, Deutschland
big Leak: Qwen4-27B architecture is being reverse-engineered from a merged kernel PR.
You might not need a MoE cluster for Qwen4-27B.
Intel’s vLLM XPU kernels just merged optimizations for the unreleased dense 27B: INT4 on 1-2 cards, MTP with 4 draft tokens, FP8 TP2, MTP, 4 residual streams/hyper-connections and QSA sparse attention kernels
The shapes are already leaking into the community.
Day-one local deploy energy is building.
That's why I love GLM 5.3 - it never gives up. Pair it with usestrix/strix and add a little DeepSeek V4.1 thinking and let it do some heavy work. – at Worms, Deutschland
Urban Mining. I'm collecting so much electronic devices to harvest RAM/CPU or microcontrollers. It's a no-brainer to find usage for electronics these days. – at Worms, Deutschland
My inner #German: I break me a branch off. – at Worms, Deutschland
SolidCraft 0.22 is out 🚁
solidcraft.io
- New Render workspace: GPU renders, studio lighting, materials
- Frames: profile library + members along sketch lines
- Holes from reference edges, thread catalogues
- 3D mouse calibration & button mapping
- AI agents connect out of the box (MCP on by default)
I am upgrading my trusty LLM Council based on github.com/karpathy/llm-coun… by @karpathy today.
Interesting find: GLM-5.3 "thinks" @grok lands the sharpest single point (anti-herd round gating, "agreement is not evidence"). – at Worms, Deutschland
ALT So many different point of views.
MINT Festival in #siliconworms - Over 1000 school kids from 8:00 to 13:00 - open to the public from 13:00 .
Visit minteinander-worms.de/mint-f… for more information. – at Worms, Deutschland
#eod fun after patchtacle days: spec-driven refactoring.
The famous GummyBears Demo - first shot was "flat" because I just pointed Qwen3.8-flash IQ4_XS model at an X Post.
After discovering the prompt here : reddit.com/r/Anthropic/comme… I instructed OpenCode to turn the prompt into specs using our #Speck-Kit fork: github.com/satwareAG/spec-ki…
Result: More Volume - Still not Opus 5.5 but not bad at all. – at Worms, Deutschland
ALT Bears gone flat.
ALT PumpItUp.
#nginx - the trusty static website server.
Just kidding, was hard to exploit. – at Worms, Deutschland
Did I really install the very first release of the Synology PHP 8.4.27 Patch for #DSM ?
ALT CVE-2026-109618 and CVE-2026-109619 patch for Synology is available.