Local LLM
Web LLM
High-performance In-browser LLM Inference Engine
★ 19.2k starsApache-2.0
View guide →TokenSpeed — a speed-of-light LLM inference engine designed for agentic workloads, with TensorRT-LLM-level performance and vLLM-level usability
TokenSpeed is an open-source local llm project by lightseekorg (lightseekorg/tokenspeed on GitHub). It has 2,473 GitHub stars and 336 forks, is released under the MIT license, and was last updated on GitHub on 15 October 2026.
TokenSpeed’s docs cover installation and requirements for your GPU stack.
Follow the launching guide to serve a model with the engine.
Use the server parameters, model recipes and parallelism guides to fit long, tool-heavy agentic sessions.
Scheduling, KV cache ownership and overlap timing are modelled as a finite-state machine tuned for agent request patterns.
A pluggable kernel registry includes one of the fastest MLA implementations on NVIDIA Blackwell.
KV resource reuse safety is enforced by the type system at compile time, not by convention.
New agentic-coding models (Qwen3.8 Flash Next, GLM 5.3 Flash) were supported at release.
Start from the source and the linked guides.
git clone https://github.com/lightseekorg/tokenspeed.gitInstallation and first run, per the official docs.
https://lightseek.org/tokenspeed/guides/getting-startedThe launching guide covers serving configuration.
https://lightseek.org/tokenspeed/guides/launchingThe TokenSpeed Team at LightSeek (lightseekorg).
Yes, MIT-licensed and open source.
No — it is a self-hosted inference engine; you bring the hardware and the model weights.
vLLM is the general-purpose usability benchmark and TensorRT-LLM the performance one. TokenSpeed targets both at once, with an architecture specialised for agentic workloads rather than chat traffic.
Clone the repo and follow the official getting-started and launching guides at lightseek.org/tokenspeed.
High-performance In-browser LLM Inference Engine
A high-throughput and memory-efficient inference and serving engine for LLMs
Offline inference engine for art.
vLLM-Omni — efficient inference for omni-modality models: image, audio, video, diffusion and policies.
Quantization, kernels, runtime and inference engine for mobiles, wearables, smart home and robots
Lightweight inference engine for running LLMs efficiently on any device.
The most-starred other Local LLM projects in the LibreAI catalogue — the realistic shortlist if TokenSpeed is not the right fit.
The simplest way to run open language models on your own machine.
User-friendly AI Interface (Supports Ollama, OpenAI API, ...)
LLM inference in C/C++
A high-throughput and memory-efficient inference and serving engine for LLMs
🤯 LobeHub is your Chief Agent Operator, organizing your agents into 7×24 operations by hiring, scheduling, and reporting on your entire AI team.
Free and open — grab the source, star the repo, and start creating.