huggingface/transformers v5.20.1: Release v5.20.1
Special surprise including the new model Beam from Reflection AI :partyingface:
ProofCode ↗
Key points
- Beam is a token-efficient, frontier-scale model focused on coding and agentic use cases.
- We broke loading models on older torch versions :cry: should be fixed now.
- Fix HunYuan-VL image assignment during batched generation (#49557) by @Arnav66692
- Fix DeepGEMM Mega MoE under expert parallelism(#49428) by @IlyasMoutawwakil
Sources (1)
- [1]huggingface/transformers v5.20.1: Release v5.20.1GitHub: huggingface/transformers · Oct 16, 01:02 PM
Special surprise including the new model Beam from Reflection AI :partying_face:
Beam is a token-efficient, frontier-scale model focused on coding and agentic use cases.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 16, 2026Welcome Beam: a 501B open model you can own, customize, and scale
- Oct 15, 2026Granite Switch 4.1: Chat, Decisions, and Speech from One Endpoint
- Oct 15, 2026huggingface/transformers v5.20.0: Release v5.20.0
- Oct 15, 2026huggingface/trl v1.16.0
- Oct 15, 2026StepFun Releases Step 5 Preview Open Weights: 600B MoE, 27B Active, 1M Context
- Oct 15, 2026Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio
