Rasmus Aagaard
@rasgaard.com
Industrial PhD Student @ Laerdal.com & DTU.dk, research in efficient machine learning, edge deployment and model compression
Embedding models keep getting cheaper - what required ~300M parameters three years ago now requires only 11M. In practice, that means that what runs on the server today could run in the browser tomorrow.
MTEB@mteb.org · 7h agoThe best embedding score available at the end of 2023 needed a 304M-parameter model. Today the same score comes from an 11M one — 28x fewer active parameters.
Me and my agents, just out of frame, are having fun in this harness, laughing
NobodyWho Chat is live on iOS & Android! 🥳 A fully offline AI assistant. No internet, no accounts, no cloud. Run #Gemma, #Ministral, #Granite, #LFM, #Qwen and more, entirely on your device. Vision and hearing support included. Fully open-source. Links in comments 👇
I guess my account is now dedicated to posting my latest dyno #bouldering
Probably my hardest climb so far. Super stoked:) Guess the grade? 🧐 #bouldering
The Danish Foundation Models (DFM) project is adapting our modular FlexOlmo architecture into a lighter-weight system that runs on commodity hardware—putting collaborative model building within reach of smaller research groups & organizations. 🧵
Got an extended abstract paper accepted at the GLOW workshop at IJCAI :) Turns out you can just delete 6 layers from the Whisper encoder and it doesn’t hurt transcription performance all that much (but you have to be pretty specific about which layers)
new post: how I develop recently using local models. the tooling is now good enough to do agentic workflows and everyone should give them a try! vickiboykis.com/2026/06/15/r...
Running local models is good now
Local agentic coding has gotten great over the past few months
vickiboykis.com
Seeing how SOTA models are evolving: becoming more restrictive in usage (decided by the company), less transparent (you cannot tell if the AI lab nerfed your model) + less private (your prompts are stored, no opt out) makes me much more interested in open models + local inference
Magtudredningen: De sociale medier udfordrer demokratiet - og politikerne er medskyldige
Magtudredningen: De sociale medier udfordrer demokratiet - og politikerne er medskyldige
politiken.dk
getting your weights quantized probably feels amazing for the larger model
new QAT quants are available now if you want to run Gemma 4 12B in only 6.7 GiB VRAM blog.google/innovation-a...
Gemma 4 QAT models: Optimizing model compression for mobile and laptop efficiency
We’re releasing Gemma 4 quantization-aware training checkpoints, reducing memory requirements and improving on-device performance.
blog.google
Very strongly recommend breaking out of the token discourse and touching grass from time to time to maintain a grip on reality
Legitimately feels like an unquantifiable vibe shift the last few weeks where the pendulum is swinging back to reasonable takes and people experimenting with model choice 🙏
Today we're announcing our #ReactNative bindings for NobodyWho! 🎉 Run #LLM models offline in your #react apps for #Android and #iOS. No internet. No API keys. No usage fees. Fully open-source. www.nobodywho.ai/posts/announ...
We've recently implemented the fastest static embedding model in the world by an insanely large margin Read on for additional info ↓
Researchers identifies "Super Weights" which are at most a handful of weights (amongst billions!). Models end up generating complete gibberish if these weights are removed, underlining the brittleness of LLMs and importance and careful consideration of outlier features. arxiv.org/pdf/2411.07191
For a recent lab meeting, I wrote up a grab bag of ways to think about your development as a researcher during a PhD: emerge-lab.github.io/papers/an-un... Sharing in case folks find it useful or have feedback!
emerge-lab.github.io
Flash-KMeans An IO-aware implementation of exact k-means that rethinks the algorithm around modern GPU bottlenecks. Flash-KMeans achieves 30x speedup over cuML and 200x speedup over FAISS. Paper: arxiv.org/abs/2603.09229 Code: github.com/svg-project/...
AI i valgkampen: Hjælper det kandidater, eller skræmmer det vælgerne væk? #dkvid #dkforsk
AI i valgkampen: Hjælper det kandidater, eller skræmmer det vælgerne væk?
Når politikere bruger kunstig intelligens, kan det svække deres troværdighed og skabe større afstand til vælgerne.
videnskab.dk
Together with fellow ddsc.io members I helped a bit with this piece on @ing.dk :)
Open source er en strategisk nødvendighed i AI: Sådan kommer vi med i kapløbet | DataTech
På tværs af it-branchen spiller open source software en fundamental rolle for både offentlige og private løsninger. Dette gør sig også gældende i AI-verdenen.Transformer-arkitekturen, grundstenen i nu...
pro.ing.dk
First blogpost of my cofounder post-Delhi on the deafening silence of public/political sphere as we come closer to another industrial revolution. anastasiastasenko.substack.com/p/the-countr...
LLM cloud inference dominates usage, but should it? Local models and accelerators have improved massively over recent years. Perfect routing to best local model "reduce energy consumption by 80.4%, compute by 77.3%, and cost by 73.8% versus cloud-only deployment" arxiv.org/pdf/2511.07885
Introducing ✨Tiny Aya✨, a family of massively multilingual small language models built to run where people actually are. Tiny Aya delivers strong multilingual performance in 70+ global languages in a 3.35B parameter model, efficient enough to run locally, even on a phone.
Excited to be in fantastic company amongst the speakers for IDA Driving AI 2026 :) ida.dk/driving-ai/t...
Due to residual structures in Transformer-models it's possible that many layers contribute very little to the downstream performance of the network, allowing for removal those redundant layers with little impact. openreview.net/pdf/72eeca23...
