XNews

Release of Open Source/Research Models (Meta CWM & HunyuanOCR)

Release of Open Source/Research Models (Meta CWM & HunyuanOCR)

Research institutions have released multiple open source/research models covering code understanding and OCR directions. - Meta FAIR has released the Code World Model (CWM) 32B, encouraging community research; CWM has shown competitive performance on various coding and mathematical benchmarks (pass@1, Math-500/AIME, etc.). - Tencent has launched HunyuanOCR, an end-to-end OCR VLM with 1B parameters, achieving leading results on multiple OCR benchmarks and real-world datasets, supporting multiple languages and long document parsing, and has been open-sourced. - Both releases provide open source resources to support research and development.

Key insights

Meta Releases CWM 32B Open Source Research Model to Promote Research in the Code Field

Meta FAIR has released the Code World Model (CWM) 32B and encourages the research community to utilize the open source weights for research. CWM has reported specific pass rates and scores on benchmarks such as SWE-bench Verified, LiveCodeBench, Math-500, and AIME, indicating its competitiveness in code understanding and mathematical reasoning.

HunyuanOCR: 1B Parameter End-to-End OCR Model Achieves Benchmark Leadership

Tencent has released HunyuanOCR, an end-to-end OCR model with approximately 1B parameters, utilizing a visual language architecture with native resolution visual encoders, MLP connectors, and lightweight language models. It covers text detection, parsing, information extraction, and VQA, performing excellently on benchmarks such as ICDAR, OCRBench, and OmniDocBench, and the model has been open-sourced.

Source posts