FreeToken System Enables RTX4060 Laptops to Run 35B Models Smoothly
机器之心 · wechat · 2026-08-22
Core Achievement
A joint team from UC Berkeley and MIT open-sourced FreeToken, an edge-native inference system designed for MoE models. It breaks hardware barriers, enabling a laptop RTX4060 to run Qwen3.6-35B (39.3 token/s) and a desktop RTX5090 to run DeepSeek-V4-Flash (284B parameters) on a single card.
Technical Highlights
- Full-Layer Double Buffering: Overlaps computation with data transfer, eliminating I/O wait bubbles.
- Bandwidth-Adaptive Scheduling: Dynamically splits tasks between GPU and CPU based on real-time PCIe bandwidth and CPU performance to maximize throughput.
- Agentic State Reuse: Sets lightweight checkpoints at token boundaries, allowing incremental recovery from the nearest anchor upon context edits, reducing Time-To-First-Token (TTFT) by 65-80%.
- Elastic GPU Memory: Automatically shrinks GPU Cache and offloads to CPU when background apps抢占 VRAM, ensuring zero-downtime service.
Pain Points Solved
- Traditional CPU-GPU offloading is slow due to PCIe limits; FreeToken achieves "interactive-grade" speeds.
- Solves I/O blocking caused by the destruction of sparsity during MoE prefill.
- Eliminates the need for 100% VRAM residency, leveraging CPU RAM and PCIe for cost-effective local deployment.
Usability
- Desktop apps for Windows/Linux and CLI are available.
- Proves that普及 LLMs relies not just on open weights but on edge-side scheduling engineering.
More from Infra
- ClickHouse tops $350M ARR; OpenAI usage up 10x in circular AI financing — iamKierraD · 2026-08-25
- Musk predicts space-based AI compute will exceed Earth's cumulative total within five years — r0ck3t23 · 2026-08-25
- OpenAI's in-house inference chip reportedly rivals GB300, NVIDIA impact seen as limited — ivan_bezdomny · 2026-08-25
- Lambda seeks input on model cards: add NVFP4 weights and base models? — TheZachMueller · 2026-08-25
- Perplexity releases research on Portable Computer on Spark — AravSrinivas · 2026-08-25
- Arav Srinivas: On-device models critical for sensitive docs with SOTA OCR — AravSrinivas · 2026-08-25