Hugging Face releases Transformers v5.7.0 with new model support and bug fixes

Hugging Face released version 5.7.0 of its Transformers library on April 28, 2025, introducing support for two new model families and addressing multiple critical bugs. The release added Laguna, Poolside's mixture-of-experts language model, and DEIMv2, a real-time object detection architecture spanning eight model sizes from X to Atto. Laguna implements a sigmoid mixture-of-experts router with per-layer head count flexibility, while DEIMv2 extends DETR capabilities with features adapted from DINOv3, achieving 57.8 AP with only 50.3M parameters at the X variant.

The update fixed several attention-mechanism bugs affecting production models. T5Gemma2 had a cross-attention cache layer type error for long inputs, Qwen3.5's gated-delta-net linear attention had incorrect cached forward behavior, and GraniteMoeHybrid crashed when no Mamba layers were present. Attention function dispatch was updated to align with latest model implementations across the codebase.

Generation improvements targeted continuous batching, which enables concurrent request handling in production deployments. Changes corrected KV deduplication and memory estimation for sequences exceeding 16,000 tokens — a critical use case for long-context applications. The release removed stale warnings about num_return_sequences that were incorrectly firing even when functionality worked correctly, reducing false alarms in production logs.

Kernel handling was strengthened with fixes for FP8 checkpoint configuration reading and error handling, particularly for quantized models like Qwen3.5-35B-A3B-FP8. Custom expert kernels registered from the Hugging Face Hub can now be properly loaded, and a rotary kernel incompatibility affecting Gemma3n and Gemma4 was resolved. An AutoTokenizer bug that selected the wrong tokenizer class, causing regressions in DeepSeek R1 and other models, was reverted.