Domestic GPU Players Form Open-Source Alliance, Attracting Core Developers from SGLang and Beyond

Chinese GPU maker Moore Threads is making a strategic push into open-source AI infrastructure. On May 14, 2026, the company hosted an SGLang × MUSA Meetup in Beijing, bringing together some of the most influential developers in the open-source AI community. The guest list read like a who's who of GitHub: BBuf (Xiaoyu Zhang), core developer of SGLang—the leading inference framework for large language models; Tang Zhengju, maintainer of TileLang, a next-generation operator programming ecosystem; Ma Teng, key contributor to Mooncake, the popular KVCache decoupling and transmission tool; Xiao Hang from the Beijing Academy of Artificial Intelligence, who works on Triton/FlagOS AI compilers; and R0CKSTAR, a prolific hardware-focused developer.

The meetup, titled "SGLang × MUSA," marked a significant milestone for domestic GPUs. MUSA is Moore Threads' unified computing architecture designed to minimize developer friction. The key message: domestic GPUs are no longer just about hardware specs—they are entering the ecosystem competition.

BBuf presented SGLang's Q2 2026 roadmap, highlighting full-stack optimizations for DeepSeek V4, including W4A16 quantization, MegaMoE acceleration, and sparse attention. A standout metric: SGLang's P/D separation architecture achieved 52.3k input tokens/s/node and 22.3k output tokens/s/node on 12 H100 nodes, costing 5x less than DeepSeek's official API.

R0CKSTAR, a Moore Threads engineer, demonstrated that SGLang on MUSA now runs nearly all mainstream large models out of the box, including DeepSeek, Tongyi Qianwen 3.5, GLM-4.5, FLUX, and Wan. Thanks to MUSA's three-layer CUDA compatibility stack (torch_musa, torchada, mthreads-ml-py), 99% of CUDA code runs with just one extra import line. As of May 12, Moore Threads has submitted 47 PRs to SGLang's mainline, with 41 already merged, achieving end-to-end support from environment setup to distributed inference.

Xiao Hang from BAAI showcased DeepSeek V4's day-zero adaptation on MUSA. Using FlagOS's Triton operator optimization and Moore Threads' SQMMA tensor acceleration engine, they reduced first-token latency by 56.7% and improved throughput by 23%. Key operator optimizations: FP8 matrix multiply averaged 8.85x speedup, sparse attention 6.01x.

Tang Zhengju demonstrated TileLang's potential: writing FlashAttention in 50 lines of Python achieving performance matching expert-written CUDA, and a GEMM operator in 15 lines matching CUTLASS while reducing code by 90%.

Ma Teng from Alibaba Cloud presented Mooncake's latest progress. Using RDMA P2P weight updates, Kimi K2's 1T model weight sync time dropped from 53s to 7.2s (7.37x acceleration). EPD three-level decoupling reduced first-token latency for multimodal models by 6-8x, and HiCache + Mooncake backend achieved over 90% cache hit ratio in multi-turn conversations.

Together, these efforts form a complete engineering chain: SGLang for inference framework, MUSA for GPU platform, FlagOS/Triton for operator optimization, TileLang for low-barrier kernel programming, and Mooncake for KVCache management and production deployment.

Moore Threads' CTO Zhang Yubo emphasized that MUSA is designed so developers don't have to learn a completely new paradigm. This philosophy is critical for upstream acceptance: minimal intrusion, maintainability, and reusability. The company is moving from maintaining a private fork to being an active contributor to mainline projects.

The implications are clear: domestic GPU ecosystem strategy is evolving from passive adapter to active co-builder. By embedding itself into the real engineering network of large model inference—including CI/CD, upstream PRs, day-zero support, and continuous optimization—Moore Threads is positioning itself as a long-term player on the global open-source stage. The era of domestic GPU hardware specs dominating the narrative is giving way to a new era of ecosystem competition.