Infinite intelligence. Local. Any Hardware. Peer-to-Peer Hyper Swarm. No cloud. No compromise. QVAC is the decentralized AI platform for humans and machines.
- Tweets 485
- Following 2
- Followers 12,315
- Likes 31
QVAC retweeted
Huge improvement
Parakeet Redux is compressed from NVIDIA's Parakeet TDT 0.6B v3, so we ran it against the original, which QVAC runs in Q8.
Both transcribed the same two hours of audio on the GPU of a MacBook Pro M5 Max.
QVAC took 22.3 seconds, 323x real time, and Redux took 44.2 seconds, 163x. Both times include loading the model.
4
3
72
16,316
Parakeet Redux is compressed from NVIDIA's Parakeet TDT 0.6B v3, so we ran it against the original, which QVAC runs in Q8.
Both transcribed the same two hours of audio on the GPU of a MacBook Pro M5 Max.
QVAC took 22.3 seconds, 323x real time, and Redux took 44.2 seconds, 163x. Both times include loading the model.
Just released Parakeet Redux!
A ternary speech-to-text model, built by compressing NVIDIA's Parakeet model from 1.2GB to 178MB.
Runs at 113x realtime on CPU, and beats the base model on the 25-language FLEURS benchmark while staying within 0.3 WER on English.
2
2
16
12,419
Hundreds of millions of people speak a language that AI barely handles, and that alone leaves them outside most of what the technology can do.
We built AfriSLM to close that gap: 19 Sub-Saharan African languages, running on a phone, open source.
The language you speak should not decide whether AI is useful to you.
CIO wrote about it.
cio.com/article/4223274/teth…
4
6
27
1,436
EuroNano translates between nine European languages, for a full weight of only 89 MB. Small enough to fit inside an app, and it runs on the phone instead of a server.
AfriNano does the same for eight African languages.
This means any app can ship multilingual translation, fully locally and for free.
networkworld.com/article/422…
1
4
21
1,720
Here are a few examples of what you can build by adding Genesis III to the training mix:
- STEM tutoring on a phone with no internet connexion
- Study tools for classrooms with no subscription budget
- A specialist model for medicine or maths
All of it is free and open source.
1
4
309
Research paper: arxiv.org/abs/2609.19513
Dataset: huggingface.co/datasets/qvac…
Hugging Face Collection: huggingface.co/collections/q…
Corpus under CC-BY-NC-4.0, the three research checkpoints under Apache-2.0.
1
4
269
QVAC SDK 0.20 is live. 🚀
We added support for:
- ABot-World: generates a walkable world from a single photo
- MiniMax-H3: create video with synchronised soundtrack
- Video models can now run on GPUs with less VRAM
- TranslatePsy-AfriSLM: translate between English & 19 African languages on-device
- ACE-Step can now generate entire songs from a single sentence & more new features
We also merged the latest changes in llama.cpp into QVAC Fabric, to ensure continuous compatibility.
Learn more in the thread below.
Complete release doc here: github.com/tetherto/qvac/rel…
4
8
105
2,186,168
Fabric has been synchronized to a newer llama.cpp base & learns to prefetch from host memory.
On devices with a dedicated GPU, dense models that don’t fully fit in GPU memory can keep some weights in system memory, with transfers now happening alongside prompt processing instead of blocking it.
On an RTX 5090 with 36 of 64 dense layers on CPU, that is 26 to 42% more prompt prefill throughput.
MoE caching is improved too: recently used experts stay on the GPU between tokens, while --fit automatically determines how much to cache.
1
1
184
Other items:
- Speech recognition adds the Nemotron model from the Parakeet family, and Parakeet runs faster across several GPUs and CPUs
- Supertonic text to speech runs faster on both CPU and GPU (read quoted article for more info)
- The asr-ggml, tts-ggml and audiogen-ggml addons now support CUDA
- The fit option sizes the expert cache itself, and can size a model from its metadata alone
- Cluster inference supports pipeline parallelism with overlapping micro batches. This maximizes the throughput, preventing any device in the cluster to be in a compute-idle state.
- MiniMax Music gains support for q4 quantization
- Prebuilds for asr-ggml, tts-ggml and audiogen-ggml are split into separate packages, so each platform installs only what it needs
- Named session caches stay warm across a restart or a stop, so the next turn skips a full re-prefill
1
240