范式智能 and 睿思智联 open-source ModelSphere, a one-stop production AI deployment platform

On 16 October, 范式智能 (stock code 6682.HK) and Beijing-based 睿思智联 launched ModelSphere (模域), a one-stop, production-grade AI deployment platform. The companies are offering it free to enterprise users, and the code is already published on GitHub.

According to the official description carried by QbitAI, the platform consolidates compute management, model deployment and runtime optimisation into a single workflow. The aim is to let companies embed open-source models into core business systems without deep work on driver compilation and hardware adaptation.

The feature list covers a graphical console with preset deployment templates, automatic elastic scaling, prefill-decode disaggregation, a three-tier KV cache, high-availability load balancing, and automatic resumption when a service stream breaks — the operational chores that dominate production AI work.

On performance, ModelSphere ships an AutoTune scheduling strategy plus hardware-aware adaptive load balancing. The vendor says peak system throughput improves by about 285% over its baseline, roughly four times the original performance; those are self-reported figures.

On hardware, the platform embeds the EngineX inference engine and can schedule Huawei Ascend, Iluvatar, Moore Threads and Hygon DCU accelerators in a unified or mixed pool, so users do not rewrite low-level code for each chip family — one of the most practical pain points in heterogeneous clusters.

The team says it has tuned support for DeepSeek-V4/V4-Flash, Qwen, Kimi, MiniMax and GLM, aiming to let enterprises deploy newly released models and reach production performance quickly, shortening the path from open weights to a live service.

The division of labour with vLLM and SGLang is described as layered: the inference engines optimise a single server's compute efficiency, while ModelSphere handles cluster-level management, including cache- and load-aware request routing, a shared KV cache pool, QoS-based autoscaling, automatic tuning from live data, monitoring and service management.

The code is released under Apache 2.0, which lowers the barrier for enterprises to try it, but the headline 285% and 4x figures are vendor-reported and will depend on the model, chip mix and traffic shape in each deployment. Independent benchmarking, not the launch post, will decide how far the claim travels.