Journal
Ce qu’on construit,
ce qu’on mesure.
Annonces, recherche du labo et coulisses de la plateforme. Avec, à chaque fois, les vrais chiffres.
À la uneAnnonce9 nov. 2026
Korollr et OxalisHeberg fusionnent pour donner Oikwos
Les modèles frugaux de Korollr et l’infrastructure d’OxalisHeberg ne font plus qu’un : voici Oikwos, le néo-cloud responsable.
Oikwos · 2 min de lecture
Annonce20 oct. 2026· 3 min
Nestor 2 : 2,8 millions de paramètres, 100 µJ par token
Notre premier modèle à tenir sur puce neuromorphique, un modèle ternaire, égale Nestor 1.1 avec sept fois moins de paramètres.
Annonce28 sept. 2026· 3 min
Introducing Versatile 1.1
Versatile is a zero-shot classification engine for text, image, audio and video. No training, no examples - you give it labels and it gives you typed answers back. Three speed tiers, from 5.5 ms, self-hostable.
Annonce28 sept. 2026· 5 min
Introducing Talos
Talos is a backend you describe instead of write: your tables, your rules in plain language, and one sentence per URL. A small model turns each sentence into a query plan. Available today on npm.
Annonce23 sept. 2026· 3 min
Introducing Nestor 1.1
Nestor 1.1 is our new small model: 12.7 MB, four effort levels, tool calling and web search. Available today in Korollr chat and through the API.
Recherche25 mai 2026· 3 min
Training a Model That Doesn't Know the Internet Exists
We trained a 7M parameter LLM exclusively on primary sources from 1924-1946. No Wikipedia. No modern text. No knowledge of what came after. Here's what we learned.
Observation23 mai 2026· 2 min
What Happens When You Let Four Models Train Overnight
Four models. One GPU. No coordination. We fixed it with a text file.
Recherche23 mai 2026· 6 min
What We Learned Trying to Reinvent Attention
We had an intuition that past tokens could declare their future relevance, replacing similarity-based attention with direct lookup. After dozens of hours of experiments, the result was clear - and not the one we expected.
Recherche20 avr. 2026· 2 min
Memory as Computation, Not Storage
We’ve been experimenting with a simple idea: what if memory in language models isn’t just a buffer, but a computational space ?
Recherche3 mars 2026· 1 min
Spiking Neural Networks for Edge Intelligence
We explore how event-driven SNNs can replace transformer attention for low-power edge inference, achieving competitive performance at a fraction of the energy cost.
Feuille de route3 mars 2026· 1 min
Towards Neuromorphic Silicon: Our 2030 Vision
Custom silicon designed natively for SNN workloads. Here's what we're building, why conventional GPUs are inadequate for brain-inspired AI, and the technical roadmap ahead.
Philosophie3 mars 2026· 1 min
The Case for Ultra-Small Models
Bigger isn't always better. We make the case for small, efficient, private AI - and why sub-billion parameter models are the future of personal computing.