6.12 billion requests, April 2025 – April 2026: the traffic changed hands from DeepSeek to Qwen.
Act · scene 1 of 10
Loading…
Loading 247,258 rows into DuckDB…
About the data
Source
A production trace of the Chutes LLM inference platform: 6,122,413,756 requests over 367 days. Day 0 of the trace is 11 April 2025 (per the source paper, arXiv:2608.13573); data and code are at github.com/HarvardMadSys/chutes_workload.
Aggregation
The public file rolls the requests up to one row per trace day × trace hour × model × endpoint function × streaming flag (1,337,094 rows), with request counts and sums with counts for input tokens, output tokens and time to first token.
This page embeds a smaller rollup, one row per day × hour × model family × request category (247,258 rows, about 4.8 MB), and computes everything in your browser with DuckDB.
Every average is a ratio of sums (for example total input tokens ÷ requests with an input-token count), so it stays correct under any filter.
Families: the five largest named families keep their own colour; the other nine (including models listed only by UUID and the long tail outside the top 200 models) are grouped as "Other families".
How to read the scenes
One bead = 10 million requests, rounded per month (largest remainders, so a month's beads add up to its rounded total). A family-month under 5 million requests can show no bead; every bead's tooltip gives the exact figure.
"Requests a day" divides a month's requests by the days present in the trace, so the partial first month (20 days) and last month (12 days) compare fairly. Partial months are marked on every month scene.
Scene titles describe all the data. With a filter on, the marks and numbers follow it; the titles do not.
Month and day scenes keep the whole year in view when you pick dates (they dim what lies outside); the KPI numbers, the time strip and the hour scenes count only your dates.
Limitations
Hour of day is the trace's own clock; the source states no time zone, so hour patterns are relative.
13 January 2026 has only 3 of 24 hours and 14 January 15 hours: incomplete logging, not a drop in demand.
Token counts exist only for text endpoints; time to first token only for streaming calls (about 92% of streaming requests report it).
The trace counts requests; it says nothing about users, prices or answer quality. Model-level detail (201 models) and the streaming split are not in this page's rollup.
Not drawn: reasoning-type models rose from 17.1% of requests in May 2025 to 44.7% in November 2025 and fell to 7.2% in March 2026.