4Paradigm and RiseUnion have open-sourced ModelSphere, a production-oriented AI deployment platform that unifies compute management, model serving and runtime tuning under the Apache 2.0 license. The code is published on GitHub at modelsphere/modelsphere, which lists an Apache-2.0 license and was pushed again on the launch day.
The companies say ModelSphere sits above engines such as vLLM and SGLang: those stacks squeeze more work from a single server, while ModelSphere handles cluster routing, a shared KV cache pool, quality-of-service autoscaling, hang recovery and online AutoTune. Built-in features listed in the joint announcement include a graphical deploy panel, elastic scale-out, prefill-decode split serving, three-level KV caching, high-availability load balancing and automatic resume after a service break.
Heterogeneous chips are a stated focus. Through an EngineX inference layer, ModelSphere is said to schedule Huawei Ascend, Iluvatar, Moore Threads and Hygon DCU resources in one fleet so teams do not rewrite low-level code for each accelerator. The project already lists adaptations for DeepSeek-V4 and V4-Flash, Qwen, Kimi, MiniMax and GLM, with a goal of production-ready configs soon after new open models appear.
On performance, the announcement claims AutoTune and hardware-aware balancing can raise peak throughput by about 285% versus an internal baseline, described as roughly a fourfold jump. That number is the companies' own claim, not an independent audit, and should be read as a marketing figure until outsiders reproduce it.
In the stack diagram the partners draw, single-node engines stay responsible for kernels and batching, while ModelSphere owns request routing that notices cache locality, a unified KV pool across replicas, SLO-driven scale decisions and operational monitoring. The pitch is that enterprises which already own mixed accelerators can stand up stable LLM APIs without rebuilding a private control plane from scratch.
The release is free for enterprise use under Apache 2.0 and is framed as infrastructure software, not a financing or earnings story.
SEE ALSO: OpenCode Adds Free DeepSeek V4.1 Flash With a 200K Context Window