QwQ
A reasoning-oriented large language model released by Alibaba as part of the Qwen series. It supports a deep-thinking mode that generates an internal reasoning process before producing a final answer, with special tokens used to delineate the reasoning chain from the final response.
Timeline (last 90 days)
MultiAlloc splits a fixed quantization budget into two parts, "per-layer precision improvement" and "per-token retention," replacing uniform compression with lightweight offline profiling, and under an equal memory budget it outperforms single-granularity baselines on Mistral, Llama, and QwQ.
Open-source models such as QwQ and DeepSeek-R1 are already able to compete on par with top closed-source commercial models on this navigation task.
All Facts (3)
开源社区对o1的复现尝试包括DeepSeek-R1和QwQ，通过强化学习训练模型的推理过程是实现类似能力的关键路径
80%UnverifiedMultiAlloc splits a fixed quantization budget into two parts, "per-layer precision improvement" and "per-token retention," replacing uniform compression with lightweight offline profiling, and under an equal memory budget it outperforms single-granularity baselines on Mistral, Llama, and QwQ.
50%UnverifiedOpen-source models such as QwQ and DeepSeek-R1 are already able to compete on par with top closed-source commercial models on this navigation task.
50%