Hugging Face AI Updates: October 17, 2026

1. Reflection AI’s Beam, a 501B Apache 2.0 MoE Model, Lands on the Hub With Day-Zero Support

Hugging Face. Reflection AI has released Beam on the Hub under Apache 2.0. It is a sparse Mixture-of-Experts model with 501B total parameters and 23B active (256 experts per layer, 8 routed per token) and a 1M-token context window, aimed at coding and agentic work. Most of its 52 layers use 512-token sliding-window attention, with 9 global layers. According to the blog post, this keeps the BF16 KV cache at about 36 GiB per sequence at 1M tokens, compared with 208 GiB if every layer were global. Reflection reports 81.6% on SWE-Bench Verified, 80.1% on Terminal Bench v2.1, and 90.6% on GPQA Diamond. Beam has five reasoning-effort levels (low through max) and ships as BF16 (about 1 TB), FP8 for Hopper and Ampere, and NVFP4 for Blackwell (about 280 GB), plus a DFlash2 speculative-decoding drafter. Day-zero support covers Transformers 5.20.1+ (with parse_response for reasoning and tool calls), vLLM, SGLang, Inference Providers (Baseten, Fireworks, Together), one-click Inference Endpoints, TRL LoRA fine-tuning, and GRPO through Miles. Source