A 428-billion-parameter open-weight model has been successfully compressed from 855GB to 128GB while maintaining its performance. This significantly smaller model achieved a score of 59.0% on the SWE-Bench Pro benchmark, narrowly surpassing GPT-5.5's score of 58.6% on the same benchmark. The development highlights advancements in model compression techniques, making powerful AI models more accessible and efficient. AI
IMPACT Demonstrates significant progress in model compression, potentially making advanced AI capabilities more accessible and efficient.
RANK_REASON The item describes a research finding about model compression and performance on a benchmark, not a direct release from a frontier lab. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
