FediScanner

#opensource

← Back
05.08.2026 18:40
admin (@admin@mastodon.raddemo.host)

How to Install #ITFlow on #Debian #VPS
This article provides a guide demonstrating how to install ITFlow on Debian VPS.

What is ITFlow?
ITFlow is a free, open-source professional services automation platform designed primarily for managed service providers and IT departments. It combines client documentation, asset tracking, password management, ticketing, invoicing, accounting, domain and SSL monitoring, ...
Continued 👉 blog.radwebhosting.com/install #letsencrypt #selfhosting #opensource #selfhosted

Show Original PostReport
05.08.2026 18:34
itguyeric (@itguyeric@mastodon.social)

Spent a while nodding along to Model Context Protocol like I knew what it meant. Then I actually sat down and figured it out:

Turns out it's something sysadmins are going to be weirdly good at, because it's the same job we've always done: build a system, expose it safely, hand over exactly the access needed and nothing more.

Take a look:
itguyeric.com/blog/what-is-mcp

Show Original PostReport
05.08.2026 18:32
p (@p@pixelfed.social)

I spent some time testing IBM Granite 4.1 3B on my secondary PC and I think I've found the sweet spot!

The model was running in Q4_K_M quantization via Ollama with the Vulkan backend on Mesa drivers.

The model fits completely into the 4 GB VRAM of my AMD Radeon RX 570, reaching around 42-46 tokens/s.

Instead of asking simple questions, I put it through a small "university exam" covering a variety of topics:

Mathematics:
• Integration by parts
• Integration by substitution
• Product rule
• Complex numbers

Physics:
• Why is the sky blue?
• Why are sunsets red?
• Fermi-Dirac distribution
• Density of states

Biology:
• Why are plants green?
• Photorespiration

Programming:
• A simple C "Hello, World!" program

My impression so far:

Surprisingly good at explaining mathematics, physics, biology and programming concepts.
Very fast and responsive thanks to the model fitting entirely into VRAM.
English responses are noticeably stronger than German ones.

There are still limitations. Creative writing with unusual constraints (for example, writing an English story while omitting all articles) is difficult for this 3B model because it has to work against the language patterns it learned during training. My Gemma 4 13B running on an RX 6700 XT follows those kinds of instructions much more reliably.

Overall, I'm genuinely impressed. For a 3B model running entirely on a nearly decade-old GPU, the combination of response quality and 42-46 tokens/s is remarkable.

Seeing a modern language model explain mathematics, physics, biology and programming at around 42-46 tokens/s on a nearly decade-old Radeon RX 570 is honestly pretty amazing. Open-source AI keeps proving that older hardware still has plenty of life left.

#LocalAI #LLM #Granite4 #IBM #OpenSource #FOSS #Linux #AMD #Vulkan #Mesa #RX570 #Ollama #SelfHosted #AI #Programming #Physics #Mathematics #Biology #Upcycling #OldHardware #EdgeAI

Show Original PostReport
05.08.2026 18:27
jurgenhaas (@jurgenhaas@fosstodon.org)

My current AI opinion is fairly simple: the model is not the thing I want to depend on.

I use #OpenCode because it gives me a stable harness I control, while models stay what they are - external services I can swap when needed. Same config, same agents, same permissions across terminal, CodeNomad, IntelliJ/PhpStorm, Zed, and #buzz

Less repeated setup. Less vendor gravity. More privacy and sovereignty.

go.lakedrops.com/opencode-my-a

#OpenSource #AI

Show Original PostReport