Qwen3.8-Max Architecture: 95B Active Parameters Delivers Outperforming Compute Efficiency
thetripathi58 · x · 2026-08-06
Discussing Alibaba's newly released Qwen3.8-Max, this post highlights its Mixture-of-Experts (MoE) architecture, which activates only 95 billion parameters out of a total of 2.4 trillion. The author notes that despite running significantly less compute per request compared to heavier models, real-world testing across finance, web development, and photorealistic rendering shows it surprisingly outperforming models with much higher compute consumption.
More from Models
- t0-alpha Released: Open-Source Foundation Model for Time-Series Forecasting — fpedregosa · 2026-08-06
- MiniMax Releases H3 Omni-Modal Model: Supports Video and Native Audio Generation — RisingSayak · 2026-08-06
- MiniMax Releases H3: 33B Open-Source DiT for Image, Video, and Audio — RisingSayak · 2026-08-06
- Google Wins Benchmarks But Loses Developers — prasenx · 2026-08-06
- OpenAI Reportedly Set to Launch Astra Next Week, Largest Pretrain Since GPT-4.5 — koltregaskes · 2026-08-06
- Google's August AI Build: 90 Reusable Agent Skills, Managed Infrastructure, New Gemini Models — rseroter · 2026-08-06