ALIBABA CLOUD · MODEL STUDIO — FRONTIER SERIES
The largest Qwen ever built. A million-token mind that reasons in long chains, writes and runs its own code, and works the tools around it — measured, not marketed.
Every flagship says it's smarter. We'd rather show the delta — 3.8 Max against the previous generation, same harness, same day, no cherry-picked seeds.
* PREVIEW FIGURES — INTERNAL EVALUATION, JULY 2026. FINAL NUMBERS SHIP WITH THE MODEL CARD.
Intelligence that stops at the answer is a parlor trick. 3.8 Max is trained end-to-end to plan, call tools, observe results, and keep going until the job is done.
Delays its first token on purpose. On hard proofs and system design it drafts, backtracks, and verifies inside a private chain of thought — up to 38k reasoning tokens before you see a word.
SELF-VERIFICATION PASS ▸ 94% CATCH RATE ON SEEDED ERRORSNot 201 separate models — one representation space where meaning transfers across scripts, idioms, and registers.
Given a repo and a goal, it navigates files, edits, runs the test suite, reads the traceback, and iterates — closing its own loop instead of handing you a diff and a prayer.
Search, browsers, interpreters, MCP servers — invoked mid-sentence with structured arguments, not bolted on with prompt glue.
Not a sliding window — a working memory. Feed it the whole codebase, the full case file, every transcript, and it keeps the thread from first token to last.
A sparse mixture-of-experts transformer, 128 layers deep. Only 8 of 64 experts fire per token — flagship capacity at a fraction of the compute. Move your cursor; scroll to pull the stack apart.
A scripted preview of the inference shell — pick a scenario or type your own line and watch the token stream land.