Phison aiDAPTIV and Intel tackle local AI's memory problem with 26B-parameter models on 16 GB RAM

New memory tech powers 26-billion-parameter models on 16 GB RAM PCs

Phison says its new memory extension technology can run a 26-billion-parameter language model on a PC with just 16 GB of RAM, potentially allowing more advanced smart software to operate locally without relying heavily on cloud infrastructure. The company unveiled the technology, called aiDAPTIV, at Computex 2026 in Taipei as part of a collaboration with Intel. The system combines Intel Core Ultra Series 3 processors with Phison's storage-based memory extension platform to support larger models and longer-running workloads on consumer PCs. As smart applications become more capable, they increasingly require more memory to handle larger models, maintain session history, and execute multi-step tasks. Many current PCs lack enough DRAM to run these workloads efficiently, forcing users to depend on cloud-based services. Phison says aiDAPTIV addresses this limitation by extending working memory beyond traditional DRAM and into high-performance NAND flash storage. The technology uses what the company calls Pascari aiDAPTIV Cache Memory to make additional memory resources available to local workloads. According to Phison, internal testing showed that a 26-billion-parameter model could run on a system equipped with 16 GB of DRAM when aiDAPTIV was enabled. The same workload required 32 GB of DRAM without the technology under identical test conditions. The company said the platform also supports runtime features such as KV cache reuse, which helps retain information from previous interactions and reduces the need to repeatedly process the same data. The collaboration with Intel is focused on enabling aiDAPTIV on Intel AI PC platforms powered by Core Ultra processors. The companies are also working on support for Intel's OpenVINO toolkit and evaluating optimized workloads for future performance demonstrations. "AI PCs are evolving into platforms for more sophisticated local AI workloads, including agentic applications and larger MoE models that place increasing demands on memory capacity and responsiveness," said KS Pua, CEO and Founder at Phison Electronics. "Through our collaboration with Intel, aiDAPTIV helps expand the necessary memory available to AI workloads on Intel AI PC platforms, allowing OEMs, developers and end users to run more capable AI applications locally while maintaining privacy and infrastructure efficiency." At Computex, the companies demonstrated a local chat interface running a mixture-of-experts model that would normally exceed the available system memory. Phison also showcased a hybrid large-language-model routing system built on OpenClaw, an open-source agent framework. The demonstration allowed larger models to run locally while using cloud-based resources only when more complex requests required additional processing. Intel said memory remains one of the primary barriers to running advanced models on client hardware. "More users and businesses want to run AI locally -- faster, more private and without the cost of sending everything to the cloud," said Jim Johnson, Senior Vice President and General Manager, Client Computing at Intel. "Our collaboration with Phison enables Intel AI PC platforms to support larger local AI workloads with simpler memory configurations, so customers can turn their own data into useful applications and real business value at a lower total cost." The announcement was made at Computex 2026 in Taipei.

Intel and Phison want to solve local AI's biggest limitation

However, the expensive Pascari SSDs cost $2,516 for 1TB, raising concerns about widespread adoption given Intel's past failures with proprietary technologies like Optane memory. Not everyone wants to run local AI on their own PCs. But if you do, there's a major problem. Most sophisticated models can't fit inside the limitations of your PC's memory and storage. Phison and Intel are working on a solution. Using Phison's aiDAPTIV solution, a 26 billion parameter AI model can be run on a laptop with 16GB of RAM, versus the 32GB of RAM it would normally require. That has two advantages: bringing local AI to more laptops and allowing more powerful laptops to load even larger models or else run separate tasks alongside AI. On a PC, AI can monopolize system resources, preventing any other work from being done. This can sometimes force a user to buy a dedicated AI PC. It's a simple productivity solution, allowing either larger AI models to be run on a PC or freeing up a laptop for other tasks. Running out of room Part of AI's hardware problem is that it has to calculate tokens, whether for something as simple as asking an LLM for a poem or a more complex set of instructions to monitor oil prices and make predictions. In either case, tokens are being generated on the fly inside the video memory. (Intel traditionally split half of a laptop's system RAM between the integrated GPU and Windows, before allowing consumers to manually adjust the allocation in August 2025.) The problem is that as a user goes on and on with an LLM, it has to remember what the original prompt instructions were as well as updates. Those can be recalculated again or stored as a reference. The problem is that AI functions are typically processed in video RAM or system RAM that's shared with the GPU. The result? Everything bogs down. Since RAM is where AI functions are computed, robbing a portion of it to "store" data lessens its effectiveness. But there's a solution and it's one you're probably already familiar with. Microsoft Word runs on your PC's CPU and uses RAM to do it, but documents are stored in the cloud or in the SSD. If Word needs a document, it asks Windows to retrieve it from your SSD. Phison does something similar. What Phison's aiDAPTIV does is use high-performance, extreme-endurance NAND flash as an AI cache, storing tokens to be recalled for later use. (Technically, the cache stores the key-value (KV) data, which grows with the context length and model size.) Normally, this would slow down the entire process. What aiDAPTIV tries to do is anticipate the model's needs, intelligently sending data back and forth between the RAM and the SSD to allow you to run larger models without impacting performance. The collaboration focuses on enabling Phison's technology on Intel AI PC platforms powered by Intel Core Ultra processors, including support for the OpenVINO toolkit, the two companies said. Together, Phison and Intel are working to demonstrate the technology for software vendors, which could eventually optimize their own apps for the technology. Of course, assuming users want to run AI locally instead of in the cloud with ChatGPT or Claude. It also assumes that users will want to run "full fat" versions of their AI models, rather than quantized models that trade accuracy for lower memory requirements and higher speed. We've gone down this road before... The aiDAPTIV concept sounds simple enough, but there's a potential catch. The joint work is being performed using Phison's Pascari AI100E family of specialized SSDs, which are designed for high endurance and sustained performance. That suggests that a successful implementation may require a laptop maker to specifically buy the Pascari SSDs. At press time, a 1TB Pascari AI100E in an M.2 2280 configuration costs $2,516 at Best Buy. Intel has gone down that road before. Optane, based on 3D XPoint memory, was an entirely new type of memory co-developed between Intel and Micron, and it behaved more like traditional memory than flash. But a lack of consumer demand forced Intel to shut down its Optane SSDs in 2021 before writing off half a billion dollars in inventory a year later. A quarter century ago, I also covered the launch, delays and eventual demise of Direct Rambus DRAM, in which memory manufacturers were asked to sign on to a partnership between Intel and Rambus over a specific type of PC memory. While memory vendors publicly agreed, privately they trashed Rambus and its licensing requirements. The lesson? A technical advantage is one thing, but being forced into a single supplier or technology is another thing altogether. We'll see how this all plays out.