Qwen3.5-27B (Dense Model) and Qwen3-Coder-Next-80B (MoE Model)

Qwen3.5-27B (Dense Model)
Released on 24 February 2026, this is the "smarter" choice for complex reasoning where every parameter is active during generation. Vertu +1

• Performance: Matches GPT-5 mini performance on the SWE-bench Verified coding benchmark (72.4%).
• Key Feature: Native multimodal capabilities (early fusion for text, images, and video).
• Hardware: Can run on consumer devices with as little as 22GB of VRAM (e.g., Mac M-series).
• Context: Supports a native context window of 256K tokens, extensible up to 1 million. YouTube +1

Qwen3-Coder-Next-80B (MoE Model)
Released on 2 February 2026, this model uses an ultra-sparse Mixture-of-Experts (MoE) architecture for extreme efficiency in coding and agentic workflows. Qwen AI +4

• Efficiency: Despite having 80B total parameters, it only activates 3B parameters per token, giving it the speed of a tiny model with the knowledge base of a large one.
• Performance: Delivers Sonnet 4.5-level coding performance and excels at long-horizon reasoning and tool usage.
• Hardware: Optimized for local deployment on hardware with 46GB to 64GB of RAM.