Useful AI models now fit in megabytes and run on phones and robots

A 22 MB vision model that does five jobs at once

Cactus Compute says its new vision model, Mirage, packs scene description, object detection, text reading, segmentation, spatial-relation identification and decision-making into a single pass, in a model with 51.3 million parameters that fits in about 22 MB using 2-bit quantization. The company designed Mirage for phones, wearables, robots and other consumer devices, and reports end-to-end processing of 74 milliseconds on a Samsung Galaxy S26 neural processing unit and 107 milliseconds on the same phone's CPU. That speed matters because it puts multi-task vision inside the latency budget of a device that never talks to the cloud. m:…7428

Mirage outputs structured JSON, including a scene graph with names, attributes and relations, plus object masks and text, and the company says it can be used with existing language models without new adapters or weight changes. In a test pairing Mirage with Qwen3-0.6B, the combination scored 31.3 out of 100 on VQA consensus credit. The model also describes images in English, German, French, Spanish, Italian, Dutch and Polish. This suggests the practical pitch is not a better score on a leaderboard but a small perception module that slots into systems developers already run. m:…7428

Teaching a small model a new skill by rewriting one matrix

A separate report describes a different kind of small-model win: a method that taught Qwen3.5-0.8B, a small open model, to classify customer-support conversations into 60 predefined scenarios across five domains, lifting accuracy from 41% to 83%. The method rewrote one weight matrix, the down-projection matrix in the model's final layer, in closed form, with no gradient steps, adapters or new parameters. The weight change was solved with ridge regression in a single step, using a dataset of 3,600 synthetic conversations, and the whole refit, from recording conversations to test numbers, takes about 45 minutes. m:…7120

There is a trade-off. The model's general-English perplexity rose from 17.4 to 19.6 after the rewrite, a roughly 12% increase, which means the edited model got better at the one job and slightly worse at general English. This suggests the technique fits teams that need a small model tuned for one narrow task on cheap hardware, and who can accept a small loss of general ability in exchange. m:…7120

Quantization that keeps bigger models close to full accuracy

In a preprint, a framework called ProQuant takes on the same problem from the other direction: instead of shrinking the model, it shrinks the numbers the model runs on. ProQuant improves W4A8 Qwen3-VL-2B-Instruct accuracy by 6.6% over existing methods and stays within 1% of full-precision models at W4A8 quantization. It does this with self-relabeling, which aligns the fine-tuning dataset distribution with the pretrained dataset distribution, and progressive block activation, which prioritizes recovery of multimodal understanding during quantization. The implication is that 2B to 8B parameter multimodal models can run on edge devices at low precision without giving up most of their accuracy. m:…2458

Optical flow on a microcontroller

In a second preprint, a hardware-aware architecture search framework called PACE-NAS optimizes dense optical flow networks for microcontrollers with limited SRAM, decoupling memory constraints from latency during the search. The resulting model, MCUFlowNet-S, runs 1.76x faster than retrained EdgeFlowNet on Ethos-U55 hardware, with a 1.6% increase in endpoint error. Replacing transposed convolutions with bilinear upsampling and convolutions cuts peak SRAM by 18% on the same hardware. Taken together, the four results point in one direction: the binding constraint on useful AI is shifting from model size to how cleverly a small model is built, taught and squeezed onto the hardware at hand. m:…9598