Qwen3.8-Max Turns 100K into 410K in Year-Long AI E-Commerce Sim
APPSO · wechat · 2026-08-04
Among the Qwen3.8-Max releases, an E-CommerceBench "long-horizon task" stands out: the model autonomously ran a simulated e-commerce store for 365 days with a 100,000 RMB initial fund, ultimately earning over 410,000 RMB through evolving negotiation strategies and precise capital allocation.
Memory Flaws in Long-Horizon Tasks
This year-long test acts as a "fault fluoroscope." In the joint MerchantBench paper by Alibaba and Zhejiang University, the previous-gen model developed an 83-day memory hallucination at day 282, mistakenly halting inventory replenishment. This highlights how models suffer from "memory distortion" during extended interactions.
The Decisive Role of Agent Frameworks
The tests also reveal the critical importance of external scaffolding. The same base model achieved 187.8% higher final net worth using the Hermes framework compared to ReAct. This proves that beyond the base model, the agent's execution and orchestration layer fundamentally dictates the ceiling of task performance.
More from coding & agent
- Overcoming Agent Brittleness: Long-Horizon Tasks Emerge as New LLM Frontier — hrishioa · 2026-08-04
- Developer Advocates for CLI to Avoid AI Tool Vendor Lock-in — yacineMTB · 2026-08-04
- Multi-Model Cross-Review: Open-Sourcing the Council Skill for Agent Evaluation — FlyFission · 2026-08-04
- Ditching Claude Code: Migrating Complex Workflows to Open-Weight Models — TheZachMueller · 2026-08-04
- Reducing LLM API Costs: How Semantic Cache Handles Repetitive Queries — qdrant_engine · 2026-08-04
- Opinion: Bash Is All AI Agents Need for Tool Use — NielsRogge · 2026-08-04