Unified memory allows the CPU and GPU to operate within a shared memory architecture; it does not mean that a graphics card has 128GB of dedicated VRAM. The operating system, model weights, context cache, and other services all consume this capacity. Up to 96GB can be allocated to the GPU, a configurable maximum rather than a guarantee of exclusive model access at all times. Compared with lower-memory devices, this creates capacity worth evaluating for larger models or multiple resident components. More memory does not automatically increase tokens generated per second, and NPU TOPS cannot be directly converted into a particular language model’s inference speed.

The relevant purchasing questions are specific: which model, which quantization level, what context length, how many concurrent users, and what response-time requirement? Even for workloads described as using a 70B or 120B model, changes in architecture, quantization, and cache settings can alter memory requirements and performance. The ability to launch a model must be assessed separately from its suitability for daily use.

How Do Private Data, Models, Storage, and Agents Work Together?
The Max version’s five-bay design provides expandable storage for local information, model files, and outputs. After access controls and any necessary indexing or retrieval are established, compatible local models can support questions and analysis. Agents can invoke tools within their permissions and save results to agreed locations. Usable capacity depends on installed drives, storage layout, and backup arrangements. Multiple bays do not mean that 100TB is included.

An individual researcher can evaluate answers grounded in their own papers and notes. An enterprise team can assess comparisons across historical projects and internal standards within access boundaries. Developers can evaluate local code and technical-document analysis. Professional conclusions still require human review, especially for legal, medical, scientific, or financial material. The greater the consequence of an answer, the more important it is to inspect original sources and omissions.

The Max version’s value extends beyond its processor. Hardware supplies compute and storage; the agent framework coordinates models, files, and tools with authorization; selected interaction channels support task submission and result review; and external models can provide an explicitly chosen supplement. Where a workflow combines local and cloud capabilities, the files and prompts that leave the device must be identified for that configuration. Storage bays alone do not establish that no data is ever uploaded.

Can Local Video Generation Justify Choosing the Max version?
We designed the Max version to support two usage directions: knowledge work and local content creation, using the same hardware configuration. Document question answering should be evaluated for the target model, answers and sources, response time, and data flows. Video generation requires assessment of asset-handling boundaries, generation time, failed runs and retries, and whether the resulting shots are usable. The final reward contents will specify which models and workflows are included.

We are preparing local video-creation workflows based on ComfyUI, LTX 2.5, and selected models. The final models and workflows may change according to licensing, hardware performance, and storage requirements. Creators who cannot readily send footage to external services, generate content relatively infrequently, and can accommodate measured processing times can consider testing the Max version once the final reward contents are announced. Those requiring repeated iterations within an afternoon should first measure the target model and shot settings; memory capacity alone does not establish production throughput. Local execution avoids the corresponding cloud-model usage charges, but electricity, drives, human review, and any external API calls still incur costs.

When Should You Choose the Max version?