Mastodawn

J'ai pris un moment pour jouer avec les données de #LLM de @betagouv

Un modèle local capable consomme 40 à 120x moins d'énergie par token qu'un frontier!!!

Pour mon cas, utiliser les modèles locaux pour le "quotidien" (rédaction, brainsto, recherches) et garder les API frontier seulement pour le dev, c'est ❗une emprunte carbone divisée par 80❗

+ la confidentialité des données :))

Note : la conso comparIA est un ordre de grandeur

#data #AI #localLLMs

Classement - compar:IA, le comparateur d'IA conversationnelles

compar:IA est un outil permettant de comparer à l’aveugle différents modèles d'IA conversationnelle pour sensibiliser aux enjeux de l'IA générative (biais, impact environmental) et constituer des jeux de données de préférence en français.

Marcus Schuler 5d ago

NVIDIA's RTX Spark brings CUDA and 128GB unified memory to mainstream Windows PCs this fall. The move reshapes local-AI hardware decisions: the question shifts from 'buy the biggest card you can afford' to 'which constraint fails first—memory, bandwidth, or model quality.' https://www.implicator.ai/nvidias-rtx-spark-splits-the-local-ai-hardware-decision-in-two/ #AI #Hardware #LocalLLMs

NVIDIA's RTX Spark Splits the Local AI GPU Decision

NVIDIA's RTX Spark drops CUDA and 128GB of unified memory into mainstream Windows PCs, sharpening a local-AI buying decision that now turns on what fails first: memory capacity, bandwidth, or model quality. What an individual and a startup should actually buy in 2026, and where to stop spending.

Implicator.ai

Moonstone2487 May 7

Mein #arbeitgeber labert grade in so nem #MicrosoftTeams Call für alle Mitarbeiter was von #digitalesouveranitat und dann soll ich MEHR mit #microsoft #github #copilot machen. Und selbstverständlich wird ALLES #ai. Sogar unsere TLD wechselt von .net auf .ai.
Wir sollen ganz explizit doch bitte #ki in die tägliche #Arbeit einbinden, der Vertrieb von KI schwärmen. Aber bloß "unsere" nutzen, wegen den Daten. Muss mich gleich mal informieren, ob #localLLMs erlaubt sind.
https://hessen.social/@Moonstone2487/116082676681677166

Hacker News May 2

Mini PC for local LLMs in 2026

https://terminalbytes.com/best-mini-pc-for-local-llm-2026/

#HackerNews #MiniPC #LocalLLMs #2026 #Technology #AI

Best mini PC for local LLMs in 2026 (Strix Halo era) | TerminalBytes

Strix Halo mini PCs doubled in price in six months. Here's what's worth buying for local LLMs in 2026, what to skip, and the 120W gotcha nobody mentions.

TerminalBytes

sammii Apr 28

Switched qwen3.5:4b from cloud to the Mac Mini and cut Spellcast API calls by ninety percent. Tinyvision tasks that once bled credits now run locally in seconds.

#SelfHosting #LocalLLMs #TinyVision #APIoptimization

Hacker News Apr 27

Running Local LLMs Offline on a Ten-Hour Flight

https://deploy.live/blog/running-local-llms-offline-on-a-ten-hour-flight/

#HackerNews #LocalLLMs #Offline #Flight #Technology #AIApplications #TravelTech

Running Local LLMs Offline on a Ten-Hour Flight

Update, 29 April 2026 This post was picked up on Hacker News and crossed 100 comments. I love that this sparks emotions. As always, there is a fair mix of reactions, which is exactly what you would expect from HN. Scroll to the bottom for extra responses to some of the comments. … Cloud Next was intense. I wrote the first version of this post at 9am my time while dozing off, and I did not expect much interest in it. Given the attention it is now getting, I gave the post the attention it deserved: corrected the model details, added more context about the setup, and responded to the most common points.

Deploy Live

Nicolas Fränkel 🇪🇺🇺🇦🇬🇪Apr 18

How to Run #LocalLLMs with #ClaudeCode

https://unsloth.ai/docs/basics/claude-code

How to Run Local LLMs with Claude Code | Unsloth Documentation

Guide to use open models with Claude Code on your local device.

Daniel S. Reichenbach Apr 13

Just a note for parallel universe me: flash attention is bad for #LocalLLMs

Liberty and Justice For All Apr 12

So I have been trying the new #Gemma4 models on my M1 macbook pro, specifically the gemma4:26b which is 17gb in size.

Obviously not the most challenging coding challenge and tasks but...

Much much faster response times than local models 6-12 months ago. Previously qwen, deepseek, and even Gemm3 simply took too long to be practical.

I find it incredible this can run on just my 5.5 year old laptop.

#ai #llm #ollama #localllms #llms

Daniel S. Reichenbach Apr 12

Just so we are clear: #LocalLLMs are an asset if trained and used well. But please be aware that many projects are pretending to be open source but their releases contain closed source components where it's not transparent what is going on.

Go to the source. Llama.cpp, PyTorch, etc.