Hugging Face Transformers v5.20.0 adds Muse Spark Open, Ling 3.0, UnlimitedOcr & unified export quantization
libraryTransformers v5.20.0 adds native support for seven new models: Meta's Muse Spark Open (816B-param multimodal MoE with 1M-token context, trained for agentic coding/tool use), InclusionAI Ling 3.0 Flash (hybrid KDA+MLA MoE), Aleph Alpha Kolibri1, MiniCPM-V 4.7 with canvas M-RoPE, Baidu UnlimitedOcr, UltraBERT encoders, and MOSS-Transcribe-Diarize. Exporters gain unified quantization (INT8/INT4) across Dynamo, ONNX, OpenVINO, and ExecuTorch. Breaking changes in cache_utils and S
an agent can now load Meta's Muse Spark Open MoE (816B/42B, 1M-token multimodal context), Ling 3.0 Flash, Kolibri1, MiniCPM-V 4.7, UnlimitedOcr, UltraBERT, and MOSS-Transcribe-Diarize natively via Hugging Face Transformers v5.20.0.
Major Transformers release adding 7 new model architectures (notably Meta Muse Spark Open MoE for agentic coding with 1M context, and UnlimitedOcr for long OCR), plus unified export-time quantization across Dynamo/ONNX/OpenVINO/ExecuTorch. Includes breaking cache_utils and SubConfigSpec changes.
pip install transformers==5.20.0from transformers import AutoModelForCausalLM, AutoTokenizer
# Load Meta's Muse Spark Open (multimodal MoE, 816B total / 42B active, 1M ctx)
model_id = "meta-llama/Muse-Spark-Open"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="bfloat16", device_map="auto")
prompt = "Refactor this repo to add async support."
out = model.generate(**tok(prompt, return_tensors="pt").to(model.device), max_new_tokens=256)
print(tok.decode(out[0], skip_special_tokens=True))
# Ling 3.0 Flash also loads natively (no trust_remote_code):
# AutoModelForCausalLM.from_pretrained("inclusionAI/Ling-3.0-Flash")