Alibaba's Qwen3.8 Max: 2.4T MoE Matches Claude Opus, Weights Coming Next Week
ArtificialAnlys · x · 2026-08-06
Alibaba has released Qwen3.8 Max, a 2.4T total parameter MoE model (activating 95B). It scores 56 on the Artificial Analysis Intelligence Index, a 10-point gain, tying Claude Opus 4.8 and sitting just behind Kimi K3.
Key Data & Details:
- Architecture & Pricing: Supports 1M context and multimodal input. API pricing drops to $2.00/$6.00 per 1M input/output tokens, with a $0.25 cache hit price.
- Cost & Inference: Despite lower token prices, cost per task surged to $1.14 (more than double its predecessor), driven by a massive increase in agentic turns (e.g., GDPval-AA turns jumped from 14 to 64).
- Capability Shifts: Shows major gains in scientific reasoning and coding, but hallucination rates rose from 23% to 40%, causing a regression in omniscience.
- Open Weights: Alibaba plans to release weights next week, marking its largest open-weights release to date.
More from Models
- Opinion: Free DeepSeek Model Handles 98% of Daily Tasks — sven_ai · 2026-08-06
- Dev Seeks Western-Hosted Platforms for GLM and Kimi Models — sgt102 · 2026-08-06
- DeepSeek API Surprises Developers with Built-in Web Search Integration — max_paperclips · 2026-08-06
- MiniMax-H3-experimental Tops Hugging Face Trending Models — Kijai · 2026-08-06
- Alexandr Wang Amazed: Muse Spark 1.2 Shows Unprecedented Capabilities — alexandr_wang · 2026-08-06
- MiniMax H3 Tops Video Generation Categories on DesignArena — airesearch12 · 2026-08-06