290B+ Models on a Gaming PC: FreeToken Engine Surges on GitHub

The open-source engine FreeToken is meant to run open MoE models of 290 billion parameters and up (“290B+”) on a gaming computer at interactive speeds. That is the promise from its developers at FlashML. The engine is free, licensed under Apache-2.0, and written in Python.

MoE (Mixture-of-Experts) means a model consists of many “experts”, and only some of them work on each token. FreeToken builds its whole design on this: it treats the GPU, CPU, system memory and the links between them as a single pool of resources. Frequently used experts stay in GPU memory, the others are loaded on demand, and part of the computation is handed to the CPU. The project includes adaptive CPU–GPU work sharing, streaming of layers during prefill, and a global LRU expert cache.

A separate part is built for agents. The engine saves checkpoints of state and KV cache, so edits to the context, such as tool calls or thinking blocks, do not force a full recomputation of the conversation. GPU memory can be reallocated between the expert cache and the context cache at runtime, without a restart or reloading of weights. The developers name DeepSeek-V4-Flash among the supported models.

The project was created on July 20, 2026. In 106 days it collected 16.4K stars and 1,605 forks, and today it is on GitHub’s daily trending list. The latest update came on November 3.

Local engines for large language models are not new: llama.cpp, one of the best known, has existed since March 2023 and runs models on ordinary computers. FreeToken, by its own description, is focused on one architecture, MoE, and on the limits of home hardware. FlashML (flashml.ai) is behind it, with a research paper and communities on Slack, Discord and WeChat. The interest in local MoE models is easy to see: DeepSeek-V3, released in December 2024, had 671 billion parameters, but only 37 billion are active for each token.