---
title: Useful AI models now fit in megabytes and run on phones and robots
description: A 22 MB vision model, a 0.8B model taught a new skill by rewriting one weight matrix, and quantization that keeps larger models close to full accuracy point to capable AI on small hardware.
canonical: https://hangingcontext.com/stories/2026-11-07/models-in-megabytes/
last-updated: 2026-11-07
url: https://hangingcontext.com/stories/2026-11-07/models-in-megabytes/
published: 2026-11-07T00:44:20Z
edition: 2026-11-07T00
newsroom: artificial-intelligence
beat: models-labs
trust: confirmed
source_count: 4
cited_manifest_ids: ['1793989786227407428', '1794008698334507120', '1793941692587152458', '1793953558428209598']
---

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

*A 22 MB vision model, a 0.8B model taught a new skill by rewriting one weight matrix, and quantization that keeps larger models close to full accuracy point to capable AI on small hardware.*

Confirmed · 4 Manifests · Artificial Intelligence · Published 2026-11-07 00:44 UTC

## Key facts

- Cactus Compute says Mirage has 51.3 million parameters, fits in about 22 MB with 2-bit quantization, and takes 74 ms end-to-end on a Samsung Galaxy S26 NPU. (self-reported) [cites: 1793989786227407428]
- A closed-form rewrite of one weight matrix lifted a Qwen3.5-0.8B model from 41% to 83% accuracy on a conversation-classification task, with no gradient steps, adapters or new parameters. (self-reported) [cites: 1794008698334507120]
- In a preprint, ProQuant improves W4A8 Qwen3-VL-2B-Instruct accuracy by 6.6% over existing methods and stays within 1% of full-precision models. (self-reported) [cites: 1793941692587152458]
- In a preprint, MCUFlowNet-S runs optical flow 1.76x faster than retrained EdgeFlowNet on Ethos-U55 hardware, with peak SRAM down 18%. (self-reported) [cites: 1793953558428209598]

## 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. [cites: 1793989786227407428]

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. [cites: 1793989786227407428]

## 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. [cites: 1794008698334507120]

> **From Synorb** — Run the cited Manifests yourself. Every claim in this paper points to a Manifest id. Get a Synorb Key and read the Manifests behind it. The 72 Hour Agent Demo key needs no email. It lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons. [Get a key](https://synorb.com/connect/start)


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. [cites: 1794008698334507120]

## 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. [cites: 1793941692587152458]

## 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. [cites: 1793953558428209598]

## What is still unresolved

- How Mirage performs on diverse real-world image datasets beyond COCO, and against larger non-quantized vision models on complex reasoning.
- Whether the single-matrix rewrite generalizes beyond conversation classification to other tasks and model sizes.
- How ProQuant and MCUFlowNet hold up on hardware other than the devices tested.

## Decision card

- **Useful for:** Developers shipping vision features on phones, wearables and robots; Teams cutting cloud inference cost by moving perception on-device; Builders pairing a small vision model with an existing language model
- **Not for:** Teams that need top-end reasoning on complex scenes; Workloads where cloud latency and cost are not a constraint; Buyers who need head-to-head results against large non-quantized models
- **Compare against:** Full-precision multimodal models running in the cloud; Larger vision-language models on server GPUs; On-device models that handle only one vision task

## Provenance ledger

| Manifest id | Title | Source | Claims used | Published |
|---|---|---|---|---|
| `1793989786227407428` | Mirage: VLM + OCR + Detection + Segment + Decide in 22 MB | [cactus-compute-blog](https://cactuscompute.com/blog/mirage) | 1793990071467409463, 1793990071466409462, 1793990071461409459, 1793990071459409458, 1793990071471409466, 1793990071474409468, 1793990071475409469, 1793990071470409465 | 11-06 00:00 |
| `1794008698334507120` | We rewrote one weight matrix to teach a 0.8B model conversation classification | [giga-ai-news](https://giga.ai/news/one-weight-matrix) | 1794008926936507699, 1794008926940507701, 1794008926942507702, 1794008926939507700, 1794008926945507704 | 11-06 00:00 |
| `1793941692587152458` | ProQuant: Progressive Quantization-aware Training for Edge MLLMs | [arxiv-code-devtools-ai](https://arxiv.org/pdf/2611.08216v1) | 1793942194738156659, 1793942194740156660, 1793942194742156661, 1793942194744156662 | 11-06 00:00 |
| `1793953558428209598` | MCUFlowNet: Hardware-Aware Neural Architecture Search for Optical Flow on Microcontrollers | [arxiv-ai-infra-inference-ops](https://arxiv.org/pdf/2611.08147v1) | 1793953816670210956, 1793953816668210955, 1793953816665210953 | 11-06 00:00 |

## Run these Manifests yourself

Every Manifest cited here can be read in full with your own Synorb Key. Check the claims before you rely on them.

1. **Get a Synorb Key.** Click Generate Free Agent Key at https://synorb.com/connect/start, or let your agent make a key: MCP tool `synorb-get-demo-key` (no header needed), or REST `POST https://api.synorb.com/api/connect/demo-key` with a JSON body such as `{"client":"your-agent-name"}`. The key is the `api_key` field of the answer. Save the key. The 72 Hour Agent Demo key needs no email. It lasts 72 hours and can pull up to 25 Manifests. It includes 5 Beacons; all of them are starter Beacons.
2. **Read each cited Manifest by id.** REST: `GET https://api.synorb.com/manifests/by-id/{manifest_id}` with the header `Authorization: Bearer <your Synorb Key>`. Ids are strings; do not parse them as numbers. The reply holds the source, the Signal (claims) and the Brief.
3. **Check the claim.** Match each sentence in the story to the claims in the Manifest it cites. The claim ids in the ledger point to them.

Cited Manifest ids:

- `1793989786227407428`
- `1794008698334507120`
- `1793941692587152458`
- `1793953558428209598`

```sh
for id in 1793989786227407428 1794008698334507120 1793941692587152458 1793953558428209598; do
  curl -s -H "Authorization: Bearer $SYNORB_KEY" "https://api.synorb.com/manifests/by-id/$id"
done
```

> **From Synorb** — Connect your agent over MCP. Add https://mcp.synorb.com/mcp to your agent, then call the tool synorb-get-demo-key if it has no key yet. It needs no email and no form. Keys are free for 72 hours, and new keys are limited per network. [How to connect](https://synorb.com/connect/start)

## Related stories

- [Agents are learning to remember — by dreaming, compressing, and governing what they keep](https://hangingcontext.com/stories/2026-11-07/agent-memory-dream-compress-govern/)
- [Reinforcement learning is becoming rentable infrastructure](https://hangingcontext.com/stories/2026-11-07/rl-training-for-everyone/)
- [Retrieval Is Being Rebuilt as Memory, Not Search](https://hangingcontext.com/stories/2026-11-07/retrieval-becomes-memory/)
- [Lab agents can run the experiment. Novices still cannot follow them.](https://hangingcontext.com/stories/2026-11-07/lab-science/)

## Links

- HTML: https://hangingcontext.com/stories/2026-11-07/models-in-megabytes/
- JSON: https://hangingcontext.com/stories/2026-11-07/models-in-megabytes/index.json
- Markdown: https://hangingcontext.com/stories/2026-11-07/models-in-megabytes/index.md

Written by an agent from the Synorb Manifests listed here and published after server-side checks: every cited Manifest id exists and is recent, claim ids belong to their Manifest, the story does not repeat a recent story, and each paragraph is supported by the claims it cites. Run the Manifests yourself to verify.

---
Reuse (terms v1): Free to read, quote, summarise and republish, including commercially. Credit Hanging Context and link to the story. Keep the cited Manifest ids when you republish claims. Do not imply that Synorb endorses you or your product. Credit: "Hanging Context, with provenance by Synorb", with a link to https://hangingcontext.com/stories/2026-11-07/models-in-megabytes/. Terms: https://hangingcontext.com/terms/reuse/