Qwen 3.8 now runs inside Agent OS, turning a 2.4 trillion parameter brain into a full worker.
A chat tab gives Qwen a mouth. Agent OS gives it a body: → Layer 1: The brain. Qwen, holding 1 million tokens of context → Layer 2: The hands.
Qwen 3.8 Max worked alone for 16 days, and next week anyone can use it for free.
Parameters are knobs in the AI's brain. This has 2.4 trillion. More than almost any AI on Earth.
Alibaba's Qwen 3.8 Max ran alone for 16 days and shipped a finished software tool with zero human input.
The receipts are public on GitHub: 265 commits. 127 pull requests. 151 issues closed.
Qwen 3.8 Max can work for 10 days straight by itself, and the weights are going open.
Not 10 minutes. Not 10 hours. Ten DAYS. It starts with an empty folder and ends with finished software.
Meta's Muse Spark 1.2 got tested against GPT 5.6 Soul, Fable 5, and Qwen 3.8.
The honest verdict: it lost almost every round. The tests: RPGs, racing games, flight sims, real builds you can click and play.
The Hermes Command Desk puts your whole AI team behind one door, and every new model is just 2 clicks away.
The scattered agent problem: You hired a brilliant assistant. Then locked them in a different building for every job. Chat in one building.
"kiana" a Qwen model checkpoint in Arena
I mentioned it to add sound effects and voice its much better than kimi k3 and GPT 5.6 Sol here SpaceX Falcon 9 launch, stage separation, boostback,…
Qwen 3.8 Max vs Fable 5.
Same screen. Same prompts. The real winner isn't who you'd guess. The 3 jobs: a full landing page, a working calculator, and a Flappy Bird game in…
Qwen 3.8 Max is free to try, so it got tested with 45 REAL builds instead of benchmarks.
3D games. A full website. A working operating system. A complete promo video. All built in hours.
Qwen 3.8 Max vs Claude Fable 5 vs GPT 5.6 Soul.
Same prompts. One shot each. Here's who actually wins. The test: Flappy Bird, Tetris, and a real landing page. One HTML file each. No do-overs.
On OSWorld-Verified, which tests real computer use
Qwen 3.8 Max: 86.1 Claude Fable 5: 85.0 GPT-5.6 Sol Max: 83.2 Gemini 3.1 Pro: 76.2 That's clicking, opening apps and completing desktop tasks.
Alibaba just released Qwen 3.8 Max, an open model that beats GPT and Gemini at using a computer.
The test is called OSWorld. It checks if an AI can click buttons and finish tasks on a real desktop like a person. Qwen 3.8 Max scored 86.1.
Qwen 3.8 Max Full COURSE 1 HOUR (Build & Automate Anything)
Kiana - Qwen New Mystery Model on Arena
This is the first test as usual a ballister test for comparison i have quote the post with output from opus 5 and sol 5.6 You can also play with it…
Using DeepSeek V4 Flash to create game demos will be ridiculously insane
I will prefer using V4 over GPT-5.6, Claude Opus 5, Kimi K3 Max or Qwen 3.8 Max, and Claude Fable 5
Turns out pretty well any frontier chatbot can build a “match 3” style game that works pretty well.
DeepSeek 4 flash, Sonnet 5, kimi k3, Qwen 3.8. All made a similar game all work pretty well.
5 major models landed on AIHubMix this week📅
Qwen 3.8 Max, Wan 3.0 Video, Qwen Image 3.0 and Pro, and Doubao Seedance 2.5.
Gemini Spark เข้าไทยแล้ว สาย Pro/Ultra ลองได้เลย 🇹🇭
สรุป AI สัปดาห์นี้ EP.14 1. Gemini Spark เปิดให้ไทยใช้แล้ว (AI Pro/Ultra ขึ้นไป) สั่งงานยาก ๆ แล้วรันบนคลาวด์ต่อเนื่องได้ 2.
Qwen 3.8 Max lists at $2/$6 per million against Kimi K3's $3/$15.
On cost per Intelligence Index task it's $1.13 against Kimi's $0.84 - it burns more tokens to finish the same job.
Qwen 3.8 Max isn’t interesting because it’s huge.
It’s interesting because it can actually DO things on a computer. A practical Agent OS setup looks like this: → Web search finds current info → Web…