阿里发布Qwen3.8-Max：2.4万亿参数MoE

Alibaba released Qwen3.8-Max, a 2.4 trillion parameter model that activates only about 95 bn parameters per token. A thread 🧵

• Blank folder to live-ready app. No step-by-step hand-holding. Full GitHub trace.
• Sparse mixture-of-experts, i.e. a small router picks a handful of experts for each token, so you pay for 95B worth of compute while the model holds 2.4T worth of stored knowledge.
• Context window is 1 million tokens, the longest single reply can run to 131,072 tokens, and the private thinking budget stretches to 262,000 tokens before it commits to an answer.
• Pricing lands at $2.00 per million input tokens and $6.00 per million output tokens, with cached reads down to $0.17 per million, so reusing a stable prompt prefix instead of resending it costs roughly 8 times less.
• On Terminal Bench 2.1, which checks whether a model can actually drive a real command line through a task end to end, it scored 86.6 against 84.6 for Opus 4.8 and 88.8 for GPT-5.6 Sol.
• It also posted 93.0 on PaperBench, a test of rebuilding a research paper's experiments in working code, and 92.6 on GPQA Diamond, a set of science questions written so that search engines do not help.

Some huge revelation from their official technical report.

• Given nothing but a research paper and some GPUs, it wrote about 7,600 lines of code over 5 days and ran 33 rounds of training to reproduce all 6 of the paper's findings.
• It was handed an empty folder and a command line tool to build, then left alone. After roughly 16 days of unattended operation the repository held 265 commits and 127 pull requests, with the model triggering its own builds, unit tests and end-to-end checks after every change.
• On a cryptographic chip design task it ran about 500 turns of edit, simulate and lay out, with no reference design to copy. Its first working circuit used 8,298 logic gates and it squeezed that down to 678, cutting physical chip area by 81% while still meeting timing at 500 MHz.
• A simulated year of running online stores: 600 suppliers, 7,000 products, and 152 fraudulent merchants hidden among them. It ended the year with a balance of 416,252 yuan from 100,000 yuan of starting capital, about 38% ahead of the next best model.

🧵 1.