Alibaba's Qwen3.8-27B debuts with hybrid attention for efficient long context

Alibaba's Tongyi Lab has released Qwen3.8-27B, a 27.78-billion-parameter multimodal model featuring a novel hybrid attention architecture. This design strategically replaces three out of every four attention layers with a linear attention mechanism called Gated DeltaNet, reserving full attention for critical layers. This approach significantly reduces memory pressure and computational cost, enabling the model to handle a native context window of 262,144 tokens, extendable to approximately one million tokens via YaRN scaling, while fitting on a single high-end consumer GPU. Additionally, Qwen3.8-27B incorporates Multi-Token Prediction for improved inference throughput and demonstrates strong performance on agentic coding benchmarks, reportedly surpassing models like Meta's Muse Glimmer-30B and Claude Opus 4.6 on several evaluations. AI

IMPACT Efficient long-context handling and improved inference throughput could accelerate adoption of large multimodal models in resource-constrained environments.

RANK_REASON Frontier-lab model release with novel architecture and benchmark claims. [lever_c_demoted from frontier_release: ic=1 ai=1.0]

Read on dev.to — LLM tag →

• Alibaba Group
• Claude Opus 4.6
• Gated DeltaNet
• Meta
• Multi Token Prediction
• Qwen3.8-27B
• Tongyi Lab
• Yarn

• model release
• infra

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →