Thank you :) Swift Models hit 2.2 million+ downloads / Early Access to New Models, Free Compute for Researchers

The Swift series of reasoning-efficient LLMs has achieved over 2.2 million downloads. This open-source release demonstrates that penalizing pathological overthinking in LLMs can reduce token usage by 58.3% and increase speed by 1.95 times without sacrificing accuracy. This is achieved through efficient reinforcement learning, rather than directly training for shorter thought processes. Swift 27B and Swift Flash Next models are available, including GSQ-RCO quants and uncensored variants.