DeepSeek-V4-Flash 284B on 5.3GB RAM: Mference Engine Ushers in the ‘SSD Expert Streaming’ Era

A developer has unveiled Mference, a groundbreaking inference engine that pushes the boundaries of Edge AI. Following the successful port of Qwen 3.6, Mference now enables the 284B-parameter DeepSeek-V4-Flash to run on a mere 5.3GB of RAM by streaming expert parameters directly from SSD storage.

• ▶ Paradigm Shift: Building on the TurboFieldfare logic, Mference leverages the sparse activation of Mixture-of-Experts (MoE) models. By keeping only the shared core and KV cache in RAM while streaming massive expert weights from the SSD on-demand, it effectively bypasses traditional memory bottlenecks.
• ▶ Performance Milestone: On an M5 Pro, the engine runs Gemma 2 26B-A4B using only ~2GB of RAM at a usable speed of 31-35 tok/s, proving that SSD-based streaming is no longer just a theoretical concept but a practical reality.

Bagua Insight

This breakthrough marks a pivotal decoupling of compute/RAM from storage in the Edge AI landscape. For years, LLM accessibility has been gatekept by the high cost of VRAM and unified memory. Mference demonstrates that through sophisticated I/O scheduling and MoE sparsity, consumer-grade SSDs can function as “virtual VRAM.” This effectively disrupts the high-margin memory strategies of hardware giants like NVIDIA. When a 284B model can run on a tablet, we have reached the inflection point for the democratization of massive-scale AI.

Actionable Advice

Hardware OEMs should prioritize high-bandwidth SSD integration (PCIe 5.0+) and direct-to-SoC data paths, as storage throughput is becoming the new benchmark for AI PCs. Developers should shift focus toward optimizing dynamic loading for MoE architectures rather than relying solely on aggressive quantization. Enterprises should re-evaluate the feasibility of deploying private, large-scale models on budget-friendly terminal hardware, significantly lowering the TCO for localized AI infrastructure.