Hugging Face AI Updates: November 21, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Thinking Machines Lab and Hugging Face Release Inkling 2 Mini, a 26B Multimodal MoE With 3B Active Parameters
Hugging Face. Hugging Face published the launch of Inkling 2 Mini together with Thinking Machines Lab. The model is a multimodal mixture-of-experts with about 26B total and 3B active parameters, a 1M-token context window, and image and text input. It routes each token to 8 of 128 experts, loops layers 6 to 17 three times, and ships with a DSpark drafter for speculative decoding. The post says it leads Gemma 4 31B, Qwen3.8-27B and Nemotron 3.5 Super VL on 32 of 48 benchmarks, and it is available in Transformers, llama.cpp (GGUF), MLX and SGLang. Source
2. Hugging Face Introduces Harness Adapters, RL-Trained LoRAs That Tune Qwen3.8-27B to a Specific Agent Harness
Hugging Face. Hugging Face described harness adapters, small LoRA adapters trained with reinforcement learning inside a specific agent harness. The team trained one adapter each for Pi, Hermes Agent and OpenCode on Qwen3.8-27B using TRL’s AsyncGRPOTrainer, OpenEnv, Harbor and 871 coding tasks. Under the Pi harness, Terminal-Bench 2.1 rose from about 39% to about 56.5% with roughly 45% fewer tool calls, and a merged super-adapter lifted mean pass@1 across six harness and benchmark pairs from about 39.5% to 49.0% with thinking off. The adapters, GGUF files, tasks and training recipe are published in a Hugging Face collection. Source