A 64GB MacBook Pro running Qwen Flash Next at roughly 24 tokens per second shows how Strata is pushing local AI beyond the limits of GPU memory alone: toward using SSDs to handle part of the computation. For consumer computers running large models, the constraint may no longer be only VRAM capacity.
What this is
Strata is a model-offloading technique that moves temporarily unused portions of a large model to an SSD and calls them back when needed. It is not a new model and does not improve the model’s intelligence. Its purpose is to reduce the hardware demands for running large models in terms of both RAM and VRAM.
The poster used Strata and other open-source research such as MoEShift to run Qwen Flash Next on an M1 Max with 64GB of RAM, reaching about 24 tokens per second—the token being the basic unit a model uses to process text. Qwen Flash Next is a relatively large open-source model, making it impractical to fit entirely in memory.
Industry view
What stands out is that users are beginning to accept the trade-off that a local model does not have to be the strongest available, as long as it is good enough. For summarization, preliminary organization and coding assistance, near-frontier-model capability may not be necessary. Keeping workloads local can also reduce cloud API costs and limit the transmission of data outside the device.
This experience, however, cannot be extrapolated directly to every machine and user. SSD offloading may introduce speed fluctuations, long-term read and write wear, and a higher probability of hardware failure. Strata and MoEShift also remain community projects without large-scale validation of their stability. A successful report from a single Reddit user demonstrates that the technical path is viable, not that the product is mature.
Impact on regular people
For enterprise IT: Teams with limited budgets may experiment with private deployments on high-memory computers first, reducing their reliance on ongoing API calls. Before adoption, however, enterprises still need to evaluate speed, concurrency, data retention and disk reliability; they should not treat results from individual enthusiasts as production-ready solutions.
For individual professionals: Tasks such as document organization, meeting notes and early-stage writing may increasingly be completed locally. In the near term, access to sufficient RAM, SSD performance and the ability to install and maintain the software will determine who benefits most.
For the consumer market: A combination of 64GB of RAM and a fast SSD may influence the local AI experience more than pursuing high-end graphics cards alone. If manufacturers optimize their software around this combination, the barrier to running large models for ordinary users will fall accordingly.