The useful question is what changes for users, developers or buyers, and whether the announcement stays industry context or becomes something people can actually use.

The assumption that large models require high-end GPUs is challenged by VKUE, an engine designed to run frontier-class models on limited hardware. By leveraging sparsity, VKUE activates only a fraction of a model’s parameters per token, reducing memory bandwidth demands. This approach allows a 34.7B-parameter model, Ourbox-35B-JGOS, to operate on an 8 GB laptop or even a GPU-less CPU system, where traditional methods would fail. The technique prioritizes accessibility over raw speed, making high-capacity reasoning models viable on modest hardware.

Ourbox-35B-JGOS, a sparse Mixture-of-Experts model from the Qwen3.5-MoE family, achieves usable performance despite its size. With 34.7B total parameters but only 3B active per token, the model runs efficiently on an 8 GB laptop or CPU-only system. Benchmark results show a 3.7× speedup over a dense model of similar size, with throughput reaching 20 tokens per second on the laptop. The model also delivers strong reasoning performance, scoring 86.4% on GPQA Diamond (maj@8) and 70.7% (greedy), validating its capabilities beyond synthetic tests.

VIDRAFT provides live demonstrations to verify VKUE’s performance across hardware tiers, from datacenter GPUs to consumer laptops. Users can input prompts and observe real-time token generation on two hardware paths simultaneously, with each panel displaying the model’s internal reasoning process. The demonstrations confirm that the same model weights function across vastly different hardware, from a $1,600 GPU to a CPU-only server, without modification. This flexibility addresses constraints in sovereign, edge, and public-sector deployments where cloud or high-end GPUs are unavailable.

VKUE is positioned as part of VIDRAFT’s efficiency-focused lineup, emphasizing ubiquity over peak performance. The engine enables frontier-class reasoning models to run on existing hardware, reducing reliance on expensive GPUs. By decoupling active parameter usage from total model size, VKUE expands deployment possibilities for organizations with limited resources, ensuring that advanced AI capabilities remain accessible regardless of hardware limitations.