DeepSeek-V3-4bit runs at >20 tokens per second and <200W using MLX on an M3 Ultra with 512GB. This might be the best and most user-friendly way to run DeepSeek-V3 on consumer hardware, possibly the most affordable too. You can finally run a GPT-4o level model locally, with possibly even better quality. #LLM #AI #ML #DeepSeek #OpenAI #GPT #OpenWeight #OpenSource https://venturebeat.com/ai/deepseek-v3-now-runs-at-20-tokens-per-second-on-mac-studio-and-thats-a-nightmare-for-openai/

This week I'm in #Montreal to present at #SANER2025 recent joint work with A. Gurioli and M. Gabrielli, on how to recognize AI-generated code. Key differences in our results wrt the state-of-the-art are: (1) a multilingual approach that supports 10 different programming languages, and (2) a fully #reproducible pipeline based on both #OpenWeight and #OpenData self-hosted LLMs. Talk info and #OpenAccess preprint at: https://conf.researchr.org/details/saner-2025/saner-2025-papers/36/Is-This-You-LLM-Recognizing-AI-written-Programs-with-Multilingual-Code-Stylometry

@fredbrooker not even the 3 trillion tokens worth of data from dolma?

Or, the more recent, 2 trillion tokens in common corpus?

https://huggingface.co/blog/Pclanglais/two-trillion-tokens-open