Europe Enters the Open Weights Model Race

Europe has just released one of the strongest open-source models in the world, and the whole thing was trained in Germany, huge win for the European AI ecosystem! A huge consortium came together to train a massive model in Deutsche Telekom's German Industrial AI Cloud in Munich on up to 512 NVIDIA B200s. This was a consortium that brings together DFKI, Fraunhofer IAIS and IIS, TU Darmstadt, Universität Würzburg, L3S, Lamarr, hessian.AI, ellamind and Merantix Momentum. The Soofi S 30B-A3B is a Mixture-of-Experts hybrid Mamba-Transformer (pretty cool to see Mamba in the wild!) 31.6B total parameters, 3.2B active per token, 52 layers of which 23 are Mamba-2, 23 MoE and only 6 GQA attention. It was pretrained on roughly on 26.68T tokens with German up-weighted to as much as 15.3% and context up to 1M tokens. Sovereign compute in Europe has been a talking point for years, and this is the first time it has produced a model I would put behind a production endpoint. Congratulations to Nicolas Flores Herr and the whole consortium.

In-house LLM Serving at Netflix

Netflix has published how they run their in-house LLM inference stack at massive scale to power use-cases across their organisation. Netflix shared how they serve LLMs behind the same Model Scoring Service that already handles the rest of their MLOps (ie XGBoost, TensorFlow and PyTorch models). They use primarily NVIDIA Triton for managing model loading, batching and GPU scheduling; they use a Java control plane above it handling versioning, autoscaling and multi-region rollout. They also moved away from TensorRT-LLM to vLLM in summer 2025, which allowed them to introduce custom architectures, hooks for custom decoding logic, easier debugging, and existing familiarity among practitioners. They also seem to have advanced rollout strategies like Red-Black deployments, as well as versioned APIs / models that keep one deployment per modelId / modelVersion pair when the schema changes. I have to say it's actually quite exciting to see that the MLOps field (and its experts) is transitioning in real time towards supporting large scale agentic systems, and this is bringing learnings from the last decade of production machine learning.

Raschka on Optimal Reasoning Models

Sebastian Raschka is taking his pareto curve analysis on model x reasoning x tasks to the next level, by breaking down how reasoning-effort settings at a model architecture level - as always there's some really great insights: It seems there are six reasoning levels exposed by GPT-5.6, which are actually trained relative to that reasoning level. Namely it seems that the effort label reaches the model as a system prompt or chat-template flag, but that label only means something if post-training taught the model to respond to it. The two main mechanisms that appear across the six open-weight models he reviews are the effort-conditioned SFT, and the mode-conditioned RLVR. This basically means that the actual per-token cost and length varies depending on the effort, so it's not just a configuration parameter the the same model generic model state. As always there are comprehensive examples, but one interesting one to point out is DeepSeek V4, which gives each of its Non-think, Think High and Think Max modes its own context window and length penalty before distilling them into one checkpoint, with a token cost that shrinks as effort rises. What is clear is that we don't yet fully understand how the model size and effort both interplay in practice,but we are starting to see more and more benchmarks that will give us the best combination to bring these together in the most efficient and prodctive way.

Mozilla's State of OSS AI

Mozilla's first State of Open Source AI report has been released! It surveyed 1.5k developers to measure the capability gap between OSS and closed models - tldr; it's gone: In 2024, open source models vs closed source still have a gap of 8.04% on Chatbot Arena; in February this year the gap has decreased to nearly zero; there is further variability but then again Kimi K3 just dropped! Inference for GPT-4-class cost fell from $20 to $0.40 per million tokens in 36 months and open weights now route roughly a third of OpenRouter tokens. 79% of developers use open models and only 51% ship them, against 63% for closed, and the gap widens rather than closes at enterprise scale (57% vs 73%). Mozilla's stack scoring puts standardization (2.83) and enterprise readiness (2.79) as the two weakest criteria across every layer, and the report argues the contest has moved to the orchestration layer above the model. MCP adoption grew from about 2M to 97M monthly SDK downloads in 16 months. From a security standpoint, 30+ CVEs reported in eight weeks and only ~21% of companies reporting mature agent governance. It is exciting to see open weights models helping the democratization of AI to avoid concentrated power within a small number of closed-model players.

Moonshot AI Kimi K3 Release

Moonshot AI has released Kimi K3, ruining the party for Anthropic's Fable and OpenAI's Sol! An Open Weights 2.8Tmodel with native vision and a 1M-token window. Impressive! The cost is currently the most surprising part, being extremely competitive with the same cost as Gemini Flash Lite (aka cheap!) but at the performance of Sonnet 5. The architecture combines Kimi Delta Attention with Attention Residuals, and pushes MoE sparsity to 16 active experts out of 896 under a Stable LatentMoE framework, which Moonshot reports gives around 2.5x better scaling efficiency than K2. Training uses quantization-aware training from the SFT stage with MXFP4 weights and MXFP8 activations, along with a balanced expert-parallel scheme using static shapes and no host synchronisation on the critical path. For production ML practitioners the deployment details are super relevant, as Moonshot recommends supernodes of 64 or more accelerators (aka probably not our home lab), and KDA breaks conventional prefix caching, so they have contributed a vLLM implementation to be released with the weights on July 27. In my opinion the self-reported kernel optimisation and chip design case studies are best treated as directional until the technical report lands, but still excited to see how this lands!

The Institute for Ethical AI & Machine Learning is a European research centre that carries out world-class research into responsible machine learning.

✉️ Email, 🐦 Twitter, 💼 Linkedin