AI Daily Digest · 2026-11-19
🔬 New AI Craft
1. Streaming routed experts makes oversized MoEs usable in limited memory
A bounded slot pool lets routed experts stream from SSD while reusing grouped_mm; prefetching avoids loading an entire expert bank per token. With specific quantized weights on a 32GB M5 and speculative decoding disabled, Qwen3.8-Flash-Next reached 13.9 tokens/s on short decoding. The Transformers interface remains under PR discussion, so this is an experimental path rather than a stable release feature.
https://huggingface.co/blog/moe-expert-streaming
2. Predict fine-tuning duration from operational overhead as well as FLOPs
Dharma-AI trained a CatBoost duration predictor using only configuration available before dispatch, including evaluation volume, checkpoint policy and per-GPU workload; queue time is excluded. Across 272 H200 SFT runs, five-fold validation gave a 1.71-hour mean absolute error versus 12.41 hours for the FLOP baseline. Other hardware needs recalibration, and the sparse runs above 60 hours remain less predictable.
https://huggingface.co/blog/Dharma-AI/predicting-fine-tuning-time
🛠️ Tools & Tips
1. Hugging Face moves language-model routing to OpenRouter with a key migration caveat
Inference Providers now routes language-model calls by price, speed and reliability, with provider failover and exclusion controls. These calls bill the Hugging Face account and no longer accept custom provider keys; image, video and audio generation retain their existing path.
https://huggingface.co/blog/openrouter
⭐ Open Source Highlights
(no items this day)
📰 Industry News
1. Anthropic prices its IPO at $83 per share, with ANT trading expected to start November 19
🚀 Major Releases
1. Gemini 4 Neon begins rolling out across paid API, business and Pro/Ultra access
2. Reson8 releases Apache-2.0 Rumble-2-Flash, a 0.5B speech recognizer for 40+ languages

