IT & Computers ⁽¹¹⁾

Follow-up to the Qwen3.8-27B article. I experimented with Qwen3.8-Flash-Next, the MoE bigger brother, on two of the machines from last time: Ulmus (one RTX 4090, Ryzen 7900) and the triple RTX 3090 server at work. Flash-Next is a MoE (125B parameters), only 6B active per token, plus a 51B n-gram table and a small MTP head. So cheap compute but needs vast amount of RAM/VRAM. The plan was to keep all of it in RAM, and let the GPU hold the dense layers, the KV cache and a cache of…

Here is my experience with Qwen3.8-27B, the first small model I have found to be actually useful for agentic tasks. It has good general knowledge, solid tool use, stays on track even after multiple context compactions and overall gets the job done. I've seen many telltale that it's been trained on some version of Opus. Such as "Excellent research" after a research session and being proud of having bypassed a harness restriction. I tested it on four very different machines: an Intel Arc Pro B70 at work (~2.5k€ machine), a…

It's the little things. May 2026, I received a CTD-8125 to replace my Mini DSP 10x10 HD and the two CT8150, in parallel, I wrote a iTunes inspired Linux music player (Sustain). Two fairly small things, but huge quality of life improvement. This one allows me : One blindspot, the CTD 8125 manual said that the amps drew 2W with network, I assumed that the network would always be up, it's not. In 2W sleep mode, it turns of networking. So I had to keep an analog Dante output to…

September 6th 2025, I had a mild stroke when I discovered that AMD did something very similar to Apple's M{n} generation chip : an APU with tons of shared memory between an excellent CPU and very capable GPU (comparable to an RTX 4000 Ada Mobile). The Ryzen AI Max+ 395 (this naming...), allows for 128GB of memory available to the GPU. This run gpt-oss-120b at 35+ token/s. Let me reinstate : It can run. a 120B model. at real-time speed. It's INSANE. Granted gps-oss-120b is an MoE, but still you…

My laptop (called Juniperus) has a 4Tb NVMe drive occupied at 75%, content is roughly distributed as such : Having lost a hard drive when I was 16 and having cried a river about it (it was a Maxtor 40Gb on a sunny summer's day)... I now have a strong backup policy. They are spread in 4 places using 3 different backup technologies. This server (called Tsuga) runs Debian and uses a 1Tb Micro SD. It is powered by a PoE hat. The aluminium enclosure only has one RJ45 sealed…

[this article is a work in progress] I'm a strong advocate for local hosting over cloud services provided by big tech companies like Google, Apple, Facebook, Amazon, and Microsoft (GAFAM). I have no trust in these corporations, and some of their decisions pose significant threats to civil rights. For instance, Apple's move to scan all your photos for illegal content (initially). The importance of data sovereignty and the risk of surveillance are critical issues. Thus, to the best of my ability, I prefer to host everything myself. At home, everything…

Until around 2007 [0], I was a big fan of powerful desktops as my daily driver. At that point my computer was liquid-cooled PowerMac G5 Quad 2.5 Ghz. This was a beast and I loved it. I then switched to a MacBook Pro 15" Penryn complemented by a NAS to compensate for lack of local storage. I kept on the laptop + NAS train for a while : more recently with availability of 4TB SSDs I went to a laptop-only setup. My previous laptop was the IdeaPad Pro 5 (5800U,…

I built a new computer to play with some machine learning. It uses what seems to be the best scoring case for airflow and temperature : Fractal Torrent Compact. The full setup is : The 4090 is capped to 65% TDP as going any further quickly hits drastrictly diminishing returns (anything above 75% should be considered OC). When capped, the card only runs at about 60° ! That thing really is a crazy object and product. During training the computer draws about 400W at the wall. Which sits perfectly in…

One of our US-48-500W died at work. Since Ubiquiti has bumped the price and shifted the range quite a lot, I decided to repair it to keep it as a spare. I first confirmed that the PSU was the culprit by powering the switch with my lab PSU. Using a meanwell PSU (RD-125-2448) providing both 24V and 48V rails ,we don't reach the original 500W of PoE power, but that's still enough for our needs (some phones, cameras, and access points). I didn't put much efforts into PSU placement in…

I replicated the results of Connor Bolton, Sara Rampazzi, Chaohao Li, Andrew Kwong, Wenyuan Xu and Kevin Fu. in "Blue Note: How intentional acoutic interference Damages availability and integrity in hard disk drives and operating systems". I set up one of my DIY Cabasse Spheres in front of a Fujitsu MJA2160BH. Doing a frequency sweep from 850Hz to 1.1Khz at around 102 to 105dB. When reaching close to 975Hz +/- 25Hz, the bitrate of the read file drops from 70MiB/s to a few KiB/s. Data was generated from /dev/random, secondary…

Components Intel i7 7700K 16 GB DDR4 512 GB NVMe 2x nvidia GTX 1080 in SLI 750 W Corsair PSU 3x 43" 4K TVs (11520 x 2160 px) ProSimu T1000 2DOF Fanatec Clubsport 2.5 Wheel base ClubSport Wheel Rim SPARCO® Rally ClubSport Porsche 918 Wheel ClubSport Steering Wheel F1® Esports V2 Fanatec Clubsport V3 Pedals Fanatec ClubSport Shifter SQ V 1.5 4.0 audio using 2x 8" Samson Resolv A8 active monitors (front) 2x 6" Samson Resolv A6 active monitors (rear) The whole system draws about 460W idle, and around 700W…