Find a model. Start a conversation. Explore what your machine can do.
Expert Streaming Engine (ESE) brings local AI to a desktop app. Browse models, chat on your own computer, and explore larger supported models by putting your SSD, system memory and NVIDIA graphics cards to work together. CPU-only packages are available too.
Start here → · See the app · What can my PC run? · Downloads
A large AI model can take more memory than your graphics card has. ESE can keep supported models' specialist weights on your SSD and move the parts in use through RAM and GPU memory. This is called expert streaming.
That opens up more model choices on the hardware you already have. The tradeoff is speed: your model, available memory and SSD determine how quickly it can answer. The streaming advantage applies to supported sparse models, also called Mixture of Experts (MoE) models. Dense models use ordinary placement.
| You want to… | ESE helps you… |
|---|---|
| Try local AI without starting at a terminal | Browse a model library and chat in ESE Studio |
| Explore models beyond your GPU's memory | Share the work across SSD, RAM and NVIDIA GPUs for supported sparse models |
| Spend less time guessing settings | Inspect a launch plan and measure configurations on your own PC |
| Keep local conversations on your machine | Chat with the local model server; benchmark sharing is optional and off by default |
Your model library, chat and tuning tools in one desktop app.
Find your next model. Search the Model hub or add a folder of models you already own. See model sizes and memory-fit recommendations before downloading.
Talk to it locally. Start a supported model, open Chat, and keep your conversation history on your computer.
Find settings worth keeping. The Config sweeper tries real launches and measures the results. Save a verified profile for next time.
See the chat and configuration screens
These are actual interface captures. The empty chat screen shows the starting state; it is not a generated conversation or performance demonstration.
- Choose your package. Use the installation guide to pick Windows, Linux CPU, or Linux NVIDIA CUDA.
- Add a model. Open Studio's Model hub, or add your existing GGUF model folder in Models. GGUF is the model-file format ESE reads.
- Start a conversation. Review the launch plan, start the model, and open Chat. Begin with a model and context size that fit your available memory.
The beginner guide walks through your first launch and explains the unfamiliar terms along the way.
Release: the v0.2.1 desktop installers include ESE Studio, the launcher and matching native runtimes. Get the packages and checksums from the versioned release page. Always use the instructions and checksums for the version you download.
| Your system | v0.2.1 package |
|---|---|
| Windows x64 | ESE-Studio_0.2.1_windows-x64-setup.exe or MSI; both include CPU and CUDA runtimes |
| Linux x86_64, CPU | ESE-Studio_0.2.1_linux-x86_64_cpu.AppImage, DEB or RPM |
| Linux x86_64, NVIDIA | The matching _cuda AppImage, DEB or RPM; compatible NVIDIA driver required |
Windows installation, CPU fallback and signed NSIS update checks pass locally. Physical Windows NVIDIA inference is deferred and remains unqualified. Windows installers lack Authenticode signatures; updater signatures are separate. Installation and update details →
An earlier engineering run loaded GPT-OSS 120B F16 — roughly 61 GiB of model weights — with about 18 GiB of combined GPU memory across two NVIDIA cards, about 47 GiB of usable system RAM, and NVMe storage. Recorded short-context generation was about 11.5 tokens per second on that machine.
That is a specific historical result, with its own model and settings. It gives a concrete example of memory beyond the GPU; it does not predict your PC's speed. See the hardware, timings and limits.
The selected experimental Flash-Next profile currently reaches about 19 tok/s in short diagnostic repeats. Its production admission remains disabled; it is not a beginner model recommendation. Long-context and extended workload checks are deferred. Current candidate scope.
- Model choice matters. An estimated memory fit is a starting point; the exact model, file variant and workload still need a real run. Experimental Qwen4Exp text inference stays disabled in production; vision and MTP on that path are unsupported.
- Local chat stays local. Model searches and downloads contact Hugging Face; update checks contact GitHub. External apps use their own services and privacy settings. Optional sharing sends sanitized benchmark summaries, never your prompts or responses.
- Start with the supported desktop path. Linux and Windows x86_64 packages are prepared. Linux CUDA checks cover the recorded NVIDIA hardware; Windows GPU inference, AMD/Intel acceleration and macOS are not qualified by this release. Check your PC.
From a source checkout:
./ese doctor
./ese build # uses CUDA when available; CPU otherwise
./ese plan /models/model.gguf # inspect the plan
./ese serve /models/model.gguf # start the local serverWindows source users can use ese.cmd; installed Windows packages include
ese.exe. See the engineering guide for build
requirements, policies, native controls, API access and measured results.
Studio guide · Profiles and tuning · Architecture · Build from source · Reference benchmarks · Community benchmarks · Changelog · Contributing
If ESE helps you explore what your PC can do, star the project or share a reproducible result. Found a problem? Open an issue with your model, hardware and steps to reproduce it.
ESE is MIT licensed and builds on ik_llama.cpp and llama.cpp, with their
original attribution retained. Bundled NVIDIA components have separate terms;
see third-party notices. Branch policy.