From #9 to #2 — then gone in a week
Over the trailing 90 days, Qwen-3.6-Max accounted for ~4% of deliberate model choices on Anuma — a respectable but unremarkable ninth place. Narrow the window to the last 30 days and the picture inverts: its share reaches ~18%, second only to GPT-5.4 and ahead of long-standing favorites like Claude Sonnet and every Gemini model.
No other model moved anywhere near as fast. As the chart above shows, most of the board is flat or drifting; Qwen's line is nearly vertical. The mirror image is Kimi-K2.5 — once the second-most-picked model — shedding almost all of its share as users migrate to its own successor and, increasingly, to Qwen.
Because as fast as it climbed, it was gone. In the last seven days Qwen-3.6-Max fell to essentially zero — it was retired from the catalog and its users routed to its successor, Qwen-3.7-Plus. The board above is a snapshot of a moment that has already passed: the entire arc, from rounding error to runner-up to sunset, played out inside a single month. That churn is the story — consumer model preference moves in weeks, not quarters, and a static benchmark would never catch it.
One in four picks is Chinese-origin
Qwen isn't an isolated story — it's the sharp edge of a broader shift. Group every model by where it was built and roughly a quarter of all deliberate picks on Anuma now go to a Chinese-origin model: Qwen, Kimi, DeepSeek, MiniMax, GLM, and Ling combined.
This tracks a pattern the broader market is starting to name: capable open models out of China, priced far below the frontier labs, winning real usage rather than just benchmark headlines. What's new here is the vantage point — this is what consumers reach for by choice, not what a leaderboard predicts they should.
Why the aggregator sees it first
A benchmark measures a model in a lab. A blind-vote arena measures it in a forced A/B. Neither captures a person with a real task, real stakes, and a free choice among dozens of models — quietly deciding a newcomer is worth switching to. Because Anuma routes across every major provider, that decision is exactly what our data records.
It's also why a shift like Qwen's shows up in our numbers weeks before it reaches a quarterly report: revealed preference is a leading indicator. The caveat is honesty about scale — this is a fast-moving, wallet-gated user base, not the whole market. But the direction is real, and it's early.
Methodology
Deliberate picks only — the auto router (the majority of messages) is resolved server-side by Anuma's model router, not a deliberate user choice, so it is excluded. Windows are the 90-day baseline and trailing 30- and 7-day windows ending July 7, 2026, each computed against that window's total explicit picks. Hundreds of raw model strings are canonicalized to their real models before counting. Origin is by model-author region, not inference location. We report usage, not quality — what people chose, on a wallet-gated, mobile-heavy user base that is directional and early, not a general-market sample.