A 35-billion-parameter MoE model that slashes memory bandwidth demands to make edge AI truly practical.
TL;DR: VIDRAFT's research lab, FINAL-Bench, released POCKET on July 23, 2026 — a 35-billion-parameter Mixture-of-Experts language model designed to run directly on iPhones and PCs without a dedicated GPU. By leveraging a sparse MoE architecture, POCKET reduces per-token memory transfer to roughly one-fifth that of a comparable dense model, making full-scale on-device inference a realistic option for everyday consumer hardware.
VIDRAFT's research division FINAL-Bench made waves in the on-device AI space on July 23, 2026, when it publicly released POCKET, a 35-billion-parameter Mixture-of-Experts (MoE) language model engineered from the ground up to run on personal devices — including iPhones and standard PCs with no discrete GPU. The release, which was covered by Japanese AI media outlet AI Chronicle, signals a meaningful step forward in the ongoing push to move large language model inference away from cloud servers and onto the hardware people already carry in their pockets.
POCKET — which stands for Personal On-device Compact Knowledge and Execution Transformer — is a 35B-parameter model built on a Mixture-of-Experts architecture. Unlike dense transformer models, MoE models activate only a subset of parameters for each token processed. This sparsity is the key engineering insight behind POCKET's hardware efficiency: while a comparable dense model requires transferring approximately 3.5 GB of data per token, POCKET brings that figure down to roughly 0.66 GB — about one-fifth as much. That difference is what makes real-time inference feasible on bandwidth-constrained devices like smartphones.
To serve the widest possible range of hardware, VIDRAFT's FINAL-Bench released POCKET in multiple quantization builds:
FINAL-Bench's official blog was quoted in the AI Chronicle report as emphasizing that "POCKET sacrifices nothing" — a claim meant to address the longstanding trade-off in on-device AI between model capability, inference speed, and hardware accessibility. According to the company, POCKET's combination of MoE sparsity and advanced quantization is what allows it to sidestep that trade-off rather than simply accepting a worse version of each dimension.
The model is released under an open-weight license, making it available for research and personal use. This follows a pattern established by FINAL-Bench with its earlier release of Aether-7B-5Attn, a fully open-source sovereign foundation model, reflecting VIDRAFT's broader commitment to building an open and privacy-respecting AI ecosystem.
The significance of POCKET sits at the intersection of two pressures reshaping the AI industry: privacy and cost. Because POCKET runs entirely on the user's own device, data never has to leave that device to reach a cloud API — a meaningful advantage for anyone handling sensitive personal, professional, or medical information. Equally important, the only ongoing cost of running POCKET is the electricity consumed by the device itself; there are no per-token API fees.
The competitive context is also worth noting. Major technology companies — including Apple with its OpenELM models and Microsoft with its Phi series — have been investing heavily in compact, on-device-capable language models. Against that backdrop, achieving practical iPhone inference with a 35-billion-parameter model, rather than a much smaller 1B–7B model, represents a notable technical milestone. The MoE architectural choice is central to this: it provides the raw capability of a large dense model while shedding the memory bandwidth requirements that would otherwise make it impossible to run on mobile silicon.
VIDRAFT's FINAL-Bench positions POCKET not as a niche research artifact but as a demonstration of a realistic, near-term path for deploying capable AI on edge devices — one that doesn't require users to buy specialized hardware or surrender their data to third-party servers.
Q: What is VIDRAFT POCKET and what makes it different from other large language models?
A: POCKET is a 35-billion-parameter Mixture-of-Experts model from VIDRAFT's FINAL-Bench lab, specifically designed for on-device inference. Its MoE architecture activates only a fraction of parameters per token, reducing memory bandwidth needs to roughly one-fifth of a comparable dense model — enabling it to run on iPhones and GPU-free PCs.
Q: Can POCKET really run on an iPhone without a cloud connection?
A: Yes. FINAL-Bench released a dedicated iPhone-optimized build that combines aggressive quantization with Apple Neural Engine (ANE) optimization, allowing the model to perform inference entirely on-device without sending data to any external server.
Q: Is POCKET free to use, and where can it be downloaded?
A: POCKET is available under an open-weight license for research and personal use. It was published by FINAL-Bench on Hugging Face, where users can access the various quantization builds suited to their hardware.
Source: AI Chronicle (일본) (2026-07-28) — original article