Particle.news

Download on the App Store

Alibaba Deprecates Hybrid Mode in Qwen3, Launches Dedicated Instruct and Thinking Models

The Qwen3-2507 update raises benchmark scores by splitting Instruct and Thinking models, offering a 256k-token context window.

Image
Image

Overview

  • Alibaba scrapped its hybrid thinking toggle after community feedback and benchmark results showed lower performance compared with separate modes.
  • The Qwen3-2507 lineup introduces distinct Instruct and Thinking variants for both 235B and 30B parameter models, each labeled with the “-2507” suffix.
  • Instruct-tuned Qwen3-235B-A22B-Instruct-2507 delivered a 2.8x lead over its April counterpart on the AIME25 math benchmark.
  • Thinking-tuned models posted 13%–54% improvements on AIME25, highlighting enhanced long-form reasoning under heavier workloads.
  • Alibaba released these variants under an Apache 2.0 license with BF16 and FP8 support, and the Qwen team has signaled continued hybrid research alongside an upcoming code-tuned 30B MoE.