FreeToken Open-Source Engine Runs 284B Models on Consumer GPUs

Researchers from Berkeley, MIT and others released FreeToken, an open-source inference engine that runs hundreds-of-billions-parameter MoE models on consumer PCs, achieving 22–25 tok/s for DeepSeek 284B on an RTX 5090; its repo has quickly gained 17.2k stars under Apache-2.0.

2026-11-15 ~ 2026-11-15 · 2 related posts

• FreeToken open-source engine runs 284B DeepSeek locally at 25 tok/s on RTX 5090 — aigclink · 2026-11-15
• FreeToken GitHub repo hits 17.2k stars for desktop-scale MoE serving — aigclink · 2026-11-15