The FlagOS Open Computing Global Challenge unites developers worldwide to push large-model inference performance across the FlagOS system software stack — from operator to end-to-end deployment. Season 2 is co-hosted with IEEE, SGLang, and OpenBMB.
The FlagOS Community brings together research institutions, chip companies, system vendors, and software ecosystem partners to build an open-source system software stack for diverse AI chips — an operator library, a cross-chip compiler, a parallel training and inference framework, and a cross-chip communication library, all built around "develop once, deploy across chips."
Season 1 closed in June 2026 across three tracks — operator development, multi-chip inference throughput, and long-context data annotation — and established the organizational foundation for what comes next. Season 2 narrows the focus to two deep tracks and widens the ecosystem: SGLang as the exclusive inference-framework partner, OpenBMB's MiniCPM5-2B as the target model, and IEEE co-hosting.
| Award | Basis | Prize (RMB) |
|---|---|---|
| Full-Scope Conquest | Most "Occupied" tasks | 10,000 / 8,000 / 5,000 |
| Breakthrough Achievement | Most "Conquered" tasks | 8,000 / 5,000 / 3,000 |
| Extreme Performance | Best single-task result | 500 |
| Award | Basis | Prize (RMB) |
|---|---|---|
| First place | Top throughput result | 30,000 |
| Second place | — | 20,000 |
| Third place | — | 10,000 |
Open to anyone, individually or in teams of up to 5. One participant identity per person for the whole competition; register only through the official site.
Pick one track. Submitting locks your track — no switching, and duplicate submissions across tracks are invalidated.
Work must be your own. Cite any third-party or open-source code you build on and make sure you have rights to use it.
Provided compute is for competition experiments only — no lending, mining, or unrelated workloads.
Follow each track's official format and submit through the designated channel before the deadline. Late or malformed submissions aren't evaluated.
No content involving security threats, discrimination, or privacy violations. Harassment results in disqualification at the organizers' discretion.
| Name | Remarks |
|---|---|
| Lin Yonghua | — |
| Bao Yungang | — |
| Yin Shouyi | — |
| Cui Huimin | — |
| — | Chip Company Representative |
| — | Model Company Representative |
| Name | Remarks |
|---|---|
| Liu Guang | — |
| Liu Hongyu | — |
| Men Chunlei | — |
| Lyu Mengsi | — |
| … | Full panel announced ahead of evaluation |