Request for access to MiniMax-M2.5-DFlash
#3
by Yandong996 - opened
Hi Z Lab team,
I have submitted an access request for z-lab/MiniMax-M2.5-DFlash and would like to kindly ask whether you could approve it.
We are currently evaluating speculative decoding for MiniMax-M2.5 in SGLang, especially for long-context agent workloads. We have tested EAGLE3-based draft models, but the acceptance length is relatively low in our real workload, so we would like to benchmark DFlash as a possible alternative.
Our intended use is internal benchmarking and engineering evaluation only:
- MiniMax-M2.5 target model serving with SGLang
- DFlash speculative decoding performance comparison
- Long-context agent scenarios with tool calls and structured outputs
- Metrics such as acceptance length, throughput, latency, and memory usage
We will follow the model terms and will not redistribute the weights or use them outside the approved scope.
Thank you for releasing this work and for supporting the community.