Qwen QwQ-32B is now live on Fireworks AI! We’re excited to bring this cutting-edge 32-billion-parameter model to our platform, as well as powering it on Hugging Face inference, offering users powerful new capabilities for complex reasoning and analytics. Here’s what makes QwQ-32B stand out: →Transformer architecture featuring Rotary Positional Embedding (RoPE), SwiGLU activations, RMSNorm, and Attention QKV bias. → Extended context length of 131,072 tokens—ideal for tackling large-scale problems. → Reinforcement Learning integration enhances its reasoning abilities, coding proficiency, and adaptability through reward-based learning. → Open-source accessibility on Hugging Face promotes wider research collaboration and innovation. Benchmark highlights: → GPQA (Graduate-Level Q&A): 65.2%, showcasing advanced scientific reasoning. → AIME (Mathematics Examination): 50.0%, impressive handling of high-level mathematical concepts. → MATH-500: Outstanding 90.6% accuracy across diverse mathematical challenges. → LiveCodeBench: 50.0%, strong performance in coding scenarios. With these impressive results, QwQ-32B is positioned to empower innovative applications and enhance productivity for data scientists, researchers, and developers. Try it out on Fireworks AI now: https://lnkd.in/dKQ_KJtH
This is a significant step forward for Fireworks AI. The extended context length and advanced reasoning capabilities in Qwen QwQ-32B could redefine how we approach data-intensive tasks. Looking forward to seeing how this impacts real-world applications Dmytro Dzhulgakov