Qwen 122B flagship and refreshed 27B in development

Alibaba's Qwen team is preparing two new open-weight releases: a 122-billion-parameter flagship and a refreshed 27B mid-tier model. No release dates, benchmark previews, or training details have been shared. The 122B would be Qwen's largest openly available checkpoint to date, while the 27B update would replace the current Qwen2.5-27B-Instruct released in September 2024.

The 122B release would push Qwen into the weight class occupied by Llama 3.1 405B and Nemotron 340B, though those models remain substantially larger. In FP16, 122B would require roughly 244GB of VRAM—fitting comfortably on four A100 80GB cards but out of reach for most consumer hardware. The 27B slot sits between Qwen's existing 14B and 32B tiers, a range where practitioners often trade off between inference cost and capability. Qwen's 72B Instruct model currently leads the open-weight pack on several reasoning benchmarks, outscoring Llama 3.1 70B and DeepSeek-V2.5 on MMLU-Pro and GPQA. Qwen releases have historically shipped with Apache 2.0 licensing and full weights on HuggingFace, and the 122B and 27B models are expected to follow the same pattern.