SGLang Project
SGLang is an open-source, high-performance serving framework for large language, vision-language, diffusion, and multimodal models. It aims to make production inference fast, scalable, low-latency, and accessible from a single GPU through distributed clusters.
At a Glance
- AI model developers and inference-platform teams
- Startups and enterprises operating production AI services
- Frontier-model companies and cloud providers
- Research laboratories and universities
- +1 more
AI Tools by SGLang Project
(2)SGLang-Omni
Open Source Omni Model Serving
SGLang
LLM and VLM Serving Framework
Discussions
No discussions yet
Be the first to start a discussion about SGLang Project
Latest News
Step 5 Preview: Day-0 Inference and RL Support with SGLang and Miles
Fish Audio Drama 3 Meets SGLang Omni: Day-0 Support and the Stage–Process–GPU Topology
Unified Radix Cache in SGLang: one tree for hybrid model prefix caching
RLinf x SGLang: Integrating Cosmos3 from Fine-Tuning to Efficient Parallel Evaluation
Products & Services
Open-source Python-embedded structured-generation language and high-performance serving runtime for LLMs and multimodal models, distributed as the sglang package and via Docker.
Built-in image and video generation engine for diffusion workloads, supporting production-scale serving, parallelism, and optimized GPU execution.
High-performance multi-stage serving runtime for omni, speech, TTS, ASR, music, and multimodal models, with OpenAI-compatible APIs and stage-specialized scheduling.
Open-source project in the SGLang GitHub organization for training speculative-decoding models and porting them to SGLang serving.
Market Position
SGLang competes with vLLM, TensorRT-LLM, Hugging Face TGI, DeepSpeed-Inference, LMDeploy, Guidance, and LMQL. It differentiates through a co-designed programming language/runtime, RadixAttention and broad prefix-cache reuse, fast constrained decoding, aggressive model/hardware day-0 support, and a unified stack spanning LLM, multimodal, diffusion, and speech serving.
Leadership
Founders
Ying Sheng
AI infrastructure and modeling researcher; the SGLang paper lists her with Stanford University, and later reporting identifies her as a key SGLang contributor and former xAI engineer and Databricks research scientist.
Ion Stoica
UC Berkeley professor and Databricks co-founder; SGLang originated in his UC Berkeley lab.
Joseph E. Gonzalez
UC Berkeley professor and machine-learning/systems researcher; co-author of the original SGLang paper.
Christos Kozyrakis
Stanford University professor and computer-systems researcher; co-author of the original SGLang paper.
Cody Hao Yu
Stanford-affiliated researcher and co-author of the original SGLang paper.
Shiyi Cao
UC Berkeley-affiliated researcher and co-author of the original SGLang paper.
Lianmin Zheng
Researcher and first-listed author of the SGLang paper; a core contributor to the serving-runtime work.
Liangsheng Yin
Researcher and co-author of the original SGLang paper.
Zhiqiang Xie
Researcher and co-author of the original SGLang paper.
Chuyue Sun
Researcher and co-author of the original SGLang paper.
Jeff Huang
Researcher and co-author of the original SGLang paper.
Executive Team
Ying Sheng
SGLang co-creator and core maintainer; later co-founder and CEO of RadixArk
Former xAI engineer and Databricks research scientist; key contributor to SGLang.
Ion Stoica
Academic originator / research sponsor
UC Berkeley professor and Databricks co-founder; SGLang originated in his lab.
Board of Directors
Founding Story
SGLang began in 2023 as an open-source research project in Ion Stoica's UC Berkeley lab. The initial vision was to make complex multi-call language-model programs easier to express and execute efficiently, combining a Python-embedded structured-generation language with a co-designed runtime that reused KV caches and accelerated constrained decoding.
Business Model
Revenue Model
SGLang is open-source Apache-2.0 software hosted by a nonprofit and is distributed for free through source, PyPI, Docker, and documentation. No SGLang-specific commercial subscription or usage-pricing model was identified; commercial managed infrastructure is associated with the separate RadixArk spinout.
Pricing Tiers
Apache-2.0 licensed software available through GitHub, PyPI, Docker, and source builds.
Target Markets
- AI model developers and inference-platform teams
- Startups and enterprises operating production AI services
- Frontier-model companies and cloud providers
- Research laboratories and universities
- Teams building agentic, RL, multimodal, speech, and generative-media systems
- Large-scale production LLM inference
- Agentic and multi-call language-model programs
- Retrieval-augmented generation and multi-turn chat
- Structured JSON/regex-constrained generation
- Reinforcement-learning rollouts and evaluation
- Vision-language, video, and multimodal inference
- Microsoft
- NVIDIA
- Oracle