Prod ML Hits 20k Stars in Github 🚀

Super excited to see that our Github repo on Prod ML just hit 20,000 stars 🚀 It's hard to believe how much this has grown since starting this list back in 2018 to map components in production ML systems; as of today it's grown massively with the support of almost 200 contributors, and over 1000 commits! It now contains hundreds of open source projects across dozens of sections; if you find that any OSS framework is missing please do contribute, as pull requests are more than appreciated! We also maintain a list on AI regulation (1.4k ⭐️), and a newer one on Production GenAI (50 ⭐️).

Sebastian Raschka LLM Year Review

Sebastian Raschka has dropped a year in review on LLMs in 2025 together with predictions for 2026, and as always there's some really insigthful thoughts, here's the summary from Sebastian himself: 1. Gold-level performance on reasoning models. 2. Qwen has overtaken Llama in popularity. 3. Mistral AI uses the DeepSeek V3 architecture. 4. Many contenders have emerged in the race for open-weight state-of-the-art models (eg Kimi, GLM, MiniMax, and Yi). 5. Cheaper / efficient architectures are priority in leading labs. 6. OpenAI released an open-weight model. 7. MCP joined the Linux Foundation and is becoming the stanard. # Predictions: 1. Consumer diffusion models for cheap. 2. Open weight community adoption of tool use. 3. RLVR will more widely expand into other domains. 4. Classical RAG will slowly fade. 5. LLM benchmark and performance progress will come from improved tooling and inference-time scaling rather than from training or the core model itself.

The State of MLOps 2025 Survey 🔥

The level of ML maturity of organisations in 2025 has slowly increased, and we have interesting insights on organisational departments/functions from our Prod ML Survey:

• In 2025 about 29% of organisations have set up a central machine learning platform team.
• However only 27% organisations have established a data platform / data engineering organisation.
• There is a significant increase with 17% organisations now establishing an AI Risk & Governance Function.
• However only 9% of organisations have established an AI inventory to keep track of all use-cases and models in the organisations.

If you want to dive deeper you can access the full results here: https://ethical.institute/state-of-ml-2025 🔥

Simon Willison 2025 Review

Simon Willison shared his 2025 year review on LLMs and as always there are some really great insights: 2025 was the year where LLMs became genuinely useful production tools thanks to RLVR, tool-use, multi-step workflows, etc. There has been a breakthrough of CLI and async coding agents (e.g., Claude Code/Codex/Gemini tooling) which is completely changing the software engineering trade, and ia also opening risks with "YOLO automation" with the "lethal trifecta" (private data + external comms + untrusted inputs). Chinese open-weight models are still rising to the top of public rankings under permissive licenses (taking over Meta's Llama. OpenAI is losing market dominance faster than ever expected, and Google Gemini has been taking marketshare at lightning speed. Local models are still improving, but arguably are still lagging behind closed source frontier models, however the gap is closing faster than also assumed. What we can be certain is that 2026 is going to be a very exciting year for the field!

OpenRouter 100 Trillion Token Analysis

OpenRouter has just dropped an in-depth analysis from 100 trillion tokens of real-world LLM usage, and has shared some really interesting findings from the ecosystem in 2025: From the data it is clear the ecosystem has shifted toward a multi-model production stack, with proprietary models still in the lead, but with open-weight models reaching roughly one-third of token volume. On the open model usage, there has been particularly rapid growth from Chinese OSS models (i.e. Qwen/DeepSeek). Contrary to the "productivity" narratives, OSS traffic is actually dominated by roleplay/interactive fiction (together with coding) which is something I did not expect to see at this magnitude, but makes sense. Nearly half of these 100 Trillion tokens involve routing to reasoning models, which supports the growth of tool calling, as well as much more context-heavy workloads driven disproportionately by programming. Cost-vs-usage shows counterintuitively that reductions on price don't seem to cause increase of usage (ie low price elasticity) - I personally think this is largely due to lack of cost-management maturity, and will most definitely shift next year. However there is still clear cost segmentation, where the market clearly still separates into groups where different users pick different models for different reasons. These are really interesting insights that do seem to reflect what we're currently seeing qualitatively; definitely will be an important space to keep a close eye on throughout 2026!

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