Hugging Face livestream: massive local AI speed-ups with speculative decoding
Hugging Face · youtube · 2026-10-14
Hugging Face's Local AI 202 livestream (Oct 19) covers speculative decoding in three talks plus Q&A: Merve Noyan (HF) introduces techniques like EAGLE, DFlash, and MTP; Eldar Kurtić (Red Hat) demos the speculators library; Maarten Grootendorst (Google DeepMind) shows how to speed up Gemma-4 with MTP drafters. Tooling spans llama.cpp, transformers, and the speculators library — useful for anyone optimizing local inference.
Related event: Hugging Face to Stream Live Session on Speculative Decoding for Local LLMs(3 posts)→
More from Infra
- Perplexity's SPACE sandbox runs millions of daily executions with sub-0.0002% error rate — inductionheads · 2026-10-15
- SkyPilot v0.14.0 ships dynamic job groups, multi-user Slurm GA and AWS Trainium support — skypilot_org · 2026-10-15
- OpenAI puts compute CTO Uday Ruddarraju in charge of infrastructure as Sachin Katti exits — rohanpaul_ai · 2026-10-14
- Running Qwen 3.8 flash with 256K context on a 12GB GPU: open-source local setup hits 30-38 tok/s — Present_Resource4022 · 2026-10-14
- CoreWeave bonds blow out to 15.1% yields as distress spreads across the curve — SumitGup · 2026-10-14
- Ben Bajarin: clean rooms and data center footprint are the two big drivers of the AI cycle — BenBajarin · 2026-10-14