LightSeek Foundation
LightSeek Foundation advances open research and open-source innovation for next-generation AI systems. It supports and coordinates collaborations across academia, industry, and the developer community to build open, efficient, trustworthy AI infrastructure and research.
At a Glance
- AI researchers and universities
- Open-source developers and maintainers
- LLM inference-engine and serving-platform operators
- AI model developers and model labs
- +2 more
AI Tools by LightSeek Foundation
(1)
TokenSpeed
Open Source LLM Inference Engine
Discussions
No discussions yet
Be the first to start a discussion about LightSeek Foundation
Latest News
Reflection's Beam at day 0: 501B MoE serving in NVFP4 on NVIDIA GB200 with TokenSpeed
TokenSpeed MLA 2 for NVIDIA Vera Rubin
685 tokens/s per user: serving Qwen3.8-Flash-Next on NVIDIA GB300 with TokenSpeed
Kimi K3 Optimization on GB300 — Part I
Products & Services
An open-source, production-oriented LLM inference engine for agentic workloads. It combines compiler-backed modeling, a high-performance scheduler, safe KV-resource reuse, a pluggable layered kernel system for heterogeneous accelerators, and SMG integration; its documentation exposes an OpenAI-compatible serving API.
A torch-native framework for disaggregated speculative-decoding training. It streams hidden states from dedicated inference GPUs to separate training workers through the Mooncake store over RDMA or TCP instead of writing them to disk.
An engine-agnostic, Rust-based LLM serving gateway that separates CPU-bound work from GPU inference through a native gRPC data plane. It supports multi-model routing, protocol-native agent APIs, tokenizer caching, reasoning/tool parsing, multimodal preprocessing, MCP orchestration, security, observability, and high availability. SMG graduated from LightSeek on July 31, 2026.
Market Position
LightSeek positions itself as a vendor-neutral nonprofit steward of open AI infrastructure rather than a commercial inference vendor. TokenSpeed is positioned against state-of-the-art engines such as TensorRT-LLM for agentic inference, while SMG's gateway-layer approach is complementary to engine/cluster projects such as NVIDIA Dynamo and llm-d. Its differentiation is open governance, cross-vendor NVIDIA/AMD support, and specialized work on agentic throughput, long-context caching, scheduling, kernels, and serving gateways.
Founding Story
LightSeek was established as a vendor-neutral public-good organization to accelerate transparent, collaborative, and inclusive AI research. Its initial vision is to connect researchers, engineers, hardware vendors, and the broader open-source community around efficient inference, model efficiency, alignment, education, and open infrastructure.
Business Model
Revenue Model
Mission-funded nonprofit model based on cash sponsorships, GPU compute support, engineering support, and grants/donations; LightSeek's sponsor page explicitly invites GPU compute, cash sponsorship, and engineering support.
Target Markets
- AI researchers and universities
- Open-source developers and maintainers
- LLM inference-engine and serving-platform operators
- AI model developers and model labs
- Cloud, GPU, and accelerator vendors
- Enterprises deploying agentic, multimodal, and long-context AI systems
- Low-latency, high-throughput LLM inference for coding agents and other agentic workloads
- Long-context and multi-turn serving with prefix/KV caching and speculative decoding
- Cross-vendor deployment on NVIDIA and AMD GPU platforms
- Disaggregated speculative-decoding model training at scale
- Multi-model, multi-engine production serving and routing
- Reasoning, function calling, multimodal inference, and MCP-enabled tool workflows
- Together AI
- Moonshot AI
- Thinking Machines Lab