yasakani
Local LLM engine for Windows and one NVIDIA RTX 5090, built for Qwen3.8-27B
Deep dive
The full breakdown - performance, comparisons, and setup
yasakaniLink to this section
yasakani is a local-AI tool - Local LLM engine for Windows and one NVIDIA RTX 5090, built for Qwen3.8-27B.
OverviewLink to this section
A local LLM engine for Windows and one NVIDIA RTX GPU (tested on the RTX 5090), built for one model: Qwen3.8-27B. Open source, Apache License 2.0. 日本語は README.ja.md。
YASAKANI runs Qwen3.8-27B on your own PC. In a comparison with eight other engines on the same PC (one RTX 5090, Qwen3.8-27B, one request at a time), its fastest model wrote answers faster than every one of them in all seven task categories, with thinking off and with thinking on. The tables and what they do and do not show are under Speed. It starts an OpenAI-compatible server (so chat front-ends and developer tools can use it), a terminal chat, and a benchmark. After the download it needs no internet connection.
It was built by one person in Japan. The engine is a purpose-built set of GPU kernels, a file format and a runtime for this one model, tuned on one card, the RTX 5090; it is not a general-purpose engine.
Both are Qwen3.8-27B, converted to the YASAKANI file format and compressed. You need only one of them.
The two quality rows are agreement with the original model's next-token prediction on 20 thinking transcripts, not task accuracy; see Quality. The file size is not the GPU memory needed to run the model.
yasakani is open-source, written primarily in Python, with 125 GitHub stars under the Apache 2.0 license. The latest release is v0.1.4 (2026-11-15).
Key capabilitiesLink to this section
From the project's documentation:
- The quality model has the smallest KL divergence (0.0122) and the highest top-1 agreement (95.24%) of the thirteen.
- Everything is kept inside the harness folder. A DeepSeek Harness you already use is not touched.
- One request is generated at a time.
- The release files are not code-signed, so Windows may warn about them (see SECURITY.md).
- The answers are not identical to the original model's (see Quality).
InstallLink to this section
A quick way to get started (always check the official docs for the latest):
pip install .How it fits a local-AI stackLink to this section
yasakani runs on your own hardware, so pair it with a model and a GPU sized to your needs. Use the VRAM calculator to pick a model that fits your card, and see what you can run for hardware guidance. Related local-AI tools in the directory:
SourcesLink to this section
- Source code & docs: yasakani-dev/yasakani
Stats from GitHub, 2026-11-16.
Frequently asked
Quick answers to common questions
What is yasakani?
yasakani is a other tool for local AI workloads. Local LLM engine for Windows and one NVIDIA RTX 5090, built for Qwen3.8-27B
Is yasakani free and open source?
Yes, yasakani has 125 GitHub stars and is licensed under Apache 2.0. You can self-host it for free on linux, windows.
What platforms does yasakani support?
yasakani runs on linux, windows.
What hardware do I need for yasakani?
The hardware requirements depend on which models you run. Check our hardware directory for compatible GPUs and systems. yasakani has 125 GitHub stars and an active community.
Does yasakani support GPU acceleration?
yasakani supports GPU acceleration via CUDA, Metal, or Vulkan depending on your platform. For the best performance, pair it with an NVIDIA RTX 4090 or 5090.
What are the best alternatives to yasakani?
Popular alternatives include other other tools in our directory. Browse our full collection at /tool for comparisons, community reviews, and benchmark data to find the right fit for your workflow.
How much does yasakani cost?
yasakani is free-open-source. It is completely free and open source to self-host.
Pairs well with
Complementary tools, models, and hardware