←
QVAC663 posts
QVAC avatar

QVAC

@qvac

Infinite intelligence. Local. Any Hardware. Peer-to-Peer Hyper Swarm. No cloud. No compromise. QVAC is the decentralized AI platform for humans and machines.

◷ Joined April 2025▧ 253 Photos and videos
QVAC avatar
QVAC@qvac·3h

An app built on it needs no server, so no one pays per image and it keeps working offline. The code is open source in our fork of stable-diffusion.cpp. Thanks to @StabilityAI for Stable Diffusion 2.1-base, IDKiro for SDXS and @madebyollin for the TAESD decoder.

QVAC avatar
QVAC@qvac·3h

The Neural Engine only runs models with fixed input sizes, so each model is exported for 512×512 images. The Stable Diffusion UNet is stored as a 6-bit palette, 622 MB instead of 1.6 GB, and both guidance passes of a step run in one call. Core ML now places 4,394 of its 4,399 operations on the Neural Engine.

QVAC avatar
QVAC@qvac·3h

With the same seed, the Neural Engine keeps the subject and composition of the GPU image. SDXS images match the GPU ones down to small details.

0:07
QVAC avatar
QVAC@qvac·3h

On the team's iPhone 16, Stable Diffusion 2.1-base runs 9.4x faster than on the GPU, and the memory it uses while generating peaks at 1.3 GB instead of 2.0 GB. SDXS, a model trained to finish in a single step, runs 16.7x faster. At a fraction of a second per image, an app can redraw the picture every time you change a word of the prompt.

0:07
QVAC avatar
QVAC@qvac·3h

Stable Diffusion spends 97% of its time in the UNet, the network that removes noise from the image step by step. We moved the UNet, with the small decoder that turns the result into pixels, from the iPhone's GPU to its Neural Engine, the part of Apple's chip built for AI math.

0:11
QVAC avatar
QVAC@qvac·3h

Our engineering team just shipped an improvement that makes image generation on iPhone about 10x faster! In the video, an iPhone 17 in airplane mode makes a Stable Diffusion 2.1-base image in 6.8 seconds, then an SDXS image in 200 milliseconds. Everything runs on the phone.

0:11
QVAC avatar
QVAC@qvac·6h

Same with AI: when the data center goes down, every app built on it stops answering. That's why we built QVAC: so anyone has access to intelligence that runs on their own phone or laptop, with no single point of failure. By default, models download peer to peer over @Holepunch_to. QVAC. Unstoppable intelligence.

Internet has lost its way. Heavy centralization within 3/5 datacenter players created predictable point of failures and easy targets. @tether built unstoppable open-source peer-to-peer communication protocols that can resist to the worst disasters. Build on @Holepunch_to 🍐 Use @keet_io 🦜

QVAC avatar
QVAC@qvac·Nov 17

The demo is open source (Apache 2.0) and MegaLoc is MIT licensed. You download the street photos once, with a free Mapillary token. Photos: Mapillary (CC BY-SA 4.0), Wikimedia Commons. Map: © OpenStreetMap contributors (ODbL). github.com/tetherto/qvac-exa…

QVAC avatar
QVAC@qvac·Nov 17

MegaLoc works anywhere you can collect photos of known places. After indexing, search runs offline and your photo stays on the device. A robot that loses GPS can locate itself from streets it has photographed, and an app can tell you which street an old family photo shows.

QVAC avatar
QVAC@qvac·Nov 17

Fan of GeoGuessr? The demo has a game mode: you and MegaLoc each drop a pin where the same street photo was taken, closest pin wins. MegaLoc won this game 3 rounds to 2, landing its winning pins 9 to 12 m off. The player won on near-identical facades and at the bookstalls.

0:49
QVAC avatar
QVAC@qvac·Nov 17

Show this model a street photo and it finds where it was taken, in a tenth of a second, offline. It's MegaLoc by @gabriberton and Carlo Masone, new in QVAC SDK 0.24. In the video it compares each photo with 10,000 geotagged street photos and lands within 15 m every time.

0:33
QVAC avatar
QVAC@qvac·Nov 16

Our new piece on @TechCrunch looks at the @Holepunch_to ecosystem and the comeback of peer-to-peer software. QVAC is part of it. Holepunch lets two devices talk directly, with no server in between. @keet_io uses it for messages, calls and file sharing. QVAC uses it to link your own devices into one private mesh, so a phone can hand heavier AI work to a stronger machine of yours. Read the article: techcrunch.com/sponsor/tethe…

QVAC avatar
QVAC@qvac·Nov 16

Other developer experience improvements: - calcTokens counts the tokens a string or a chat history with tools takes for the loaded model, without running inference - Fabric moves to llama.cpp b11739, which brings faster flash-attention prompt reading on Metal - Audio generation can run ACE-Step and its language model on separate devices, and returns output in chunks - GLM 5.3 Flash is available in the SDK and the AI SDK provider

QVAC avatar
QVAC@qvac·Nov 16

Transcription gained speaker attribution and more detail: - Whisper returns token timestamps and confidence, the detected language and speaker turns - Parakeet adds voice activity tuning, prewarming and device selection, with richer streaming results and lower streaming overhead - MOSS adds device selection, progress updates and detailed statistics

QVAC avatar
QVAC@qvac·Nov 16

Qwen3.5 can now run partly on the NPU in @AMDRyzen AI chips (Strix Point, Strix Halo and Krackan), so the GPU stays free for another model. It ships switched off, so you opt into it. On Qwen3.5-0.8B with a supported quantization, the NPU writes nearly all of the answer. On larger models it only reads the prompt, and the CPU writes the answer.

QVAC avatar
QVAC@qvac·Nov 16

You can now start a new chat from another chat's cache. A shared prefix, such as a system prompt, tools or documents, is processed once, and branching a chat becomes a cheap copy. The source cache stays unchanged, and the SDK refuses one made with a different model, system prompt or tool list.