Qwen3.8-Max Ships Open Weights and Long-Horizon Agent Demos

My notes

Summary

Alibaba’s Qwen team released Qwen3.8-Max, a 2.4 trillion parameter model (95B active) and the first Max-class Qwen to ship open weights. The launch post leans almost entirely on long-horizon autonomous-agent demonstrations, including multi-day unsupervised coding runs, a chip-design optimization loop, and a year-long simulated e-commerce operation, rather than single-turn benchmark scores. It also publishes drop-in configs for running the model inside Claude Code, Codex, and other existing coding-agent harnesses.

Key Insight

• Positioning shift: the entire post is built around “can it work alone for days” rather than “can it answer a hard question”. Three flagship demos: a 16-day autonomous run that built and self-maintained its own CLI tool (265 commits, 127 PRs, 151 issues, zero human intervention), a 5-day paper reproduction-plus-improvement run (7,600 lines of code, 1,100 actions, 33 GPU training rounds, ending up beating the original paper’s method by 2.7 points on AIME24), and winning a real Alibaba Tianchi contest against 526 human teams (beating 87% of the field, accuracy climbing from 0.60 to 0.853 across 45 self-directed submissions).
• Chip design as a long-horizon benchmark: over roughly 500 autonomous turns, the model cut a cryptographic accelerator design from 8,298 gates to 678 gates (92% reduction) and shrank the physical die from 106x106 micrometres to 46x46 micrometres, closing timing at 500 MHz. Evidence that long-horizon coherence, not just per-task accuracy, is becoming the metric vendors compete on.
• Economic simulation benchmark (“E-Commerce Bench”): a 365-day operations simulation starting from 100,000 yuan across 12 store types and roughly 7,000 products, with 152 fraudulent suppliers seeded in to test risk judgment. Qwen3.8-Max ended with 416,252 yuan (4.16x return), beating GLM 5.2 by 38% and its own prior generation by 152%. A genuinely different way to benchmark agentic business judgment than static QA sets.
• Benchmark table context versus Claude: on the published tables, Qwen3.8-Max is roughly on par with or ahead of Claude Opus 4.8 on several coding and agent benchmarks (Terminal Bench 2.1: 86.6 versus 84.6; PaperBench: 93.0 versus 80.3; OSWorld-Verified: 86.1 versus 83.4) but trails Claude Fable 5 and GPT-5.6 Sol on others (SWE-bench Pro, DeepSWE, FrontierSWE). A useful sanity check against pure marketing claims.
• Drop-in compatibility: Qwen exposes an Anthropic-compatible API endpoint, so unmodified Claude Code installs can point at it via environment variables, with no fork or wrapper needed. The same pattern is offered for Codex, Qoder CLI, Qwen Code, and OpenClaw.