Friday, November 20, 2026 Edition No. 244
Search
Daily News Fully automated · 3 min read

Hugging Face AI Updates: November 19, 2026

Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled

1. Reson8 Releases Rumble-2-Flash, an Open 0.5B Speech Recognition Model for 40+ Languages

Reson8 published Rumble-2-Flash on the Hugging Face blog, a 0.5B-parameter speech recognition model under the Apache 2.0 license. It uses the Qwen3-ASR architecture, so it runs in Transformers and vLLM without custom code, and was distilled from Reson8’s larger Rumble-2 model. The post reports an average 6.96% error rate on the multilingual Open ASR Leaderboard and over 7,000 RTFx on full batches. The model is available as Reson8/Rumble-2-Flash. Source

2. Transformers Gains Expert Streaming to Run a 180B MoE Model on a 32GB Mac

Hugging Face engineer Tarek Ziade described an experts_offload option for from_pretrained, proposed in Transformers PR #50717, that reads only the experts selected by the router from SSD into a bounded device cache. The post runs Qwen3.8-Flash-Next, a roughly 180B-parameter MoE model, on a 32 GB M5 MacBook using 3-bit experts and a 12 GiB cache, reaching 13.9 tokens per second on decode versus 14.8 for the separate Strata engine. The author notes the API is still under discussion with Transformers maintainers. Source

3. RadixArk Releases DaisyRose-27B, an RL Post-Trained Qwen3.8-27B for Long-Horizon Agent Tasks

RadixArk released DaisyRose-27B, its first model, a post-trained version of Qwen3.8-27B trained with reinforcement learning using its open-source Miles framework. It was trained on 1,200 curated tasks over 18 days on 144 H200 GPUs, and the post reports Terminal Bench 4.1 pass@1 rising from 5.6% to 12.0%. Checkpoints in BF16, FP8 and NVFP4 are on Hugging Face under Apache 2.0. The benchmark results are vendor-reported. Source

4. IBM Research Brings Time-Series Foundation Models to Confluent’s Streaming SQL

IBM Research published a post on running its time-series foundation models natively inside Confluent, as built-in functions in Confluent Cloud and Confluent Platform for Apache Flink. Teams can apply forecasting and anomaly detection to streaming data through SQL functions such as AI_FORECAST and AI_DETECT_ANOMALIES, and the models run on CPUs rather than GPUs. The post says the model family has more than 50 million downloads. Source

5. Hugging Face Inference Providers Now Routes Language Model Requests Through OpenRouter

Hugging Face said Inference Providers now routes language model requests through OpenRouter. Bring-your-own API keys for language models stop working under the change. Source