Qwen3.8-Max-0902 tested: 2.4T params, 1M context, and it built three full apps
JaynitMakwana · x · 2026-09-05
Blogger JaynitMakwana tested Alibaba's newly released Qwen3.8-Max-0902, billed as Alibaba's most powerful model yet: 2.4T parameters, a 1M-token context window, and extra training for coding and agent workflows, with stronger results on complex software engineering, research and long-horizon tasks.
He ran three demanding prompts:
- A full personal finance dashboard (income/expense tracking, budgets, savings goals, recurring transactions, interactive charts, dark mode, responsive design), asking the model to test every interaction and fix what breaks
- A playable 2D tower defense game (multiple enemy types, tower placement, upgrades, waves, sound, increasing difficulty) with self-testing of the game loop
- A full 2D space combat game (enemy waves, power-ups, homing missiles, multi-stage boss battle)
His takeaway: coding and agent capability is where the model stands out — it can find and fix broken mechanics on its own.
Related event: Hands-on: Alibaba's Qwen3.8-Max-0902 with 2.4T Parameters and 1M Context(2 posts)→
More from coding & agent
- M3E Canvas: Design Your App Visually, Export It as a Prompt for Coding Agents — mhdfaran · 2026-09-05
- GPT-6 in Codex still needs human steering on hard problems, researcher reports — DimitrisPapail · 2026-09-05
- More AI-built procedural scenes: first-person Sedona red rock walk with zero assets — prasenx · 2026-09-05
- Procedural 3D jungle built with Claude and Cursor: zero external art assets — prasenx · 2026-09-05
- GPT-6 Astra guide ships an official prompt to curb the model's excessive test-running habit — JeremyNguyenPhD · 2026-09-05
- Armin Ronacher: LLMs gave me the tenacity to hack a CarPlay dongle's firmware — mitsuhiko · 2026-09-05