Mobius Labs GmbH is a company focused on revolutionizing the work with technical images through AI solutions, particularly in the area of object and keyword recognition, pixel location detection, and ML algorithms on edge devices.

Computer vision has made major leaps in the last few years. Mobius Labs are going a step further: it’s no longer just Computer Vision, it’s Superhuman Vision. 
Superhuman Vision means instantly transforming thousands of images and videos into insight, experience, and advantage. Our artificial intelligence-powered technology supercharges visual searches by understanding everything about an image or a video, and automatically creating descriptive metadata for the same.
The SDK can run on cloud, local servers and edge devices (like laptops and smartphones) providing the ability to instantly understand and analyze images, video footage, and real-life scenarios. Fast, customizable and trainable with very little data, Mobius Labs' AI transforms image and video archives and makes search and tagging problems a thing of the past.
We believe that Vision is a power. Our goal is to add Superhuman Vision to any application, device or process, thus giving you unassailable competitive advantage.
To get in touch, drop us an email at marketing@mobiuslabs.com or visit our website. 


Mobius Labs GmbH is receiving additional funding by the ProFIT program of the Investment Bank of Berlin. The goal of the ProFIT project “Superhuman Vision 2.0 for every application - no code, customizable, on-premise  AI solutions ” is to revolutionize the work with technical images. (f.e.)
This project is co-financed by the European Fund for Regional Development (EFRE).
In the ProFIT project, we are exploring models that can recognize various objects and keywords in images and can also detect and segment these objects into specific pixel locations. Furthermore, we are investigating the application of ML algorithms on edge devices, including satellites, mobile phones, and tablets. Additionally, we will explore the combination of multiple modalities, such as audio embedded in videos and the language extracted from that audio.

APEX Ventures

Atlantic Labs

Lunar Ventures

Space Capital

Edward Lando

Martin Ott

Ventech

Charlie Songhurst

Marcus Ruebsam

Sridhar Ramaswamy

Mobius Labs

Excited about the potential of extreme quantization! Check out this 2-bit Mixtral quantized model running smoothly on a standard RTX 4090 GPU. 🙌 Takes only 13 GB in VRAM

HQQ's new meta-data offloading unlocks true n-bit storage on the GPU. This means massive VRAM savings and the ability to experiment with even larger models on consumer hardware. 🚀

Get the model at https://lnkd.in/efW8VCUM

Excited to announce the release of the new version of HQQ, now featuring CUDA dequantization kernels. It’s live at https://lnkd.in/e6YuWgPJ and boasts a performance improvement of approximately 2-3x over the native PyTorch compilation process.

ps : A shoutout to Jeremy Howard for his excellent tutorial demystifying CUDA kernels for us at https://lnkd.in/ecSrVzap

Releasing Aanaphi2 LLM instruct model to open source, a significant breakthrough in the 3-billion-parameter model class that has secured the top position overall in the open LLM leaderboard. Overall, Aanaphi2 achieves an average score of 63.28, edging out Microsoft's Phi-2's 61.33 and marking a new benchmark in the field. The model is available at https://lnkd.in/eSd-ba6t

And here is the official confirmation that we are ranked number 1 among 3 billion-parameter language models in the Open LLM Benchmarks. 

Huge shout out to Hicham Badri . It is always an immense pleasure to work with you and see you work!

Today, we are thrilled to unveil the Aanaphi2 instruct model, a significant breakthrough in the 3-billion-parameter model class that has secured the top position overall. Demonstrating superior performance in five out of the seven benchmarks, the Aanaphi2 model stands as a testament to its advanced capabilities and innovation. The model is available in Hugging Face from https://lnkd.in/eJVnQt_5 

In direct comparison with the current leaders in this class, the Microsoft Phi-2 models, the Aanaphi2 model shows significant performance enhancements. Overall, Aanaphi2 achieves an average score of 63.89, edging out Phi-2's 61.33 and marking a new benchmark in the field.

Our journey is propelled by a collaborative spirit of innovation, standing on the shoulders of pioneering researchers and the open-source community. Their willingness to share their work has allowed us to push the boundaries of what's possible.
We remain committed to this legacy, continually developing and releasing more cutting-edge models to contribute to the collective advancement of our field.

Watch out for more OSS drops of vision, audio and language models, hopefully SOTA

Dropping soon from Mobius Labs: Aana-phi2:  Leader in the open LLM leaderboard, 3 billion parameter category

Here is Hicham Badri  delivering a presentation on our latest work in quantizing large models. HQQ stands as one of the foremost quantization techniques globally, distinguished by its ability to operate without calibration data and its exceptional speed. For more insights, I encourage you to watch the video.

https://lnkd.in/d5u6bmTZ

I will be giving a talk about HQQ quantization this Wednesday 31st, feel free to register for the event!

We are releasing new 2-bit Mixtral models. These ones use a mixed HQQ 4-bit/2-bit configuration, resulting in a significantly improved model (ppl 4.69 vs. 5.90) with a negligible 0.20 GB VRAM increase.

Base: https://lnkd.in/dquBykfa
Instruct: https://lnkd.in/daDmBZ7Z

Shout-out to Artem Eliseev and Denis Mazur for suggesting this idea.

We are releasing the 4-bit and 2-bit quantized versions of the new Mixtral-8x7B : https://lnkd.in/dyw4x6Ga For the same memory usage ~26GB, Mixtral outperforms Llama2-70B (wiki2 ppl 3.79 vs. 4.13). The 2-bit models can run on a single 24GB GPU.

COM:MOBIUSLABS

Mobius Labs GmbH

Analysis Not Available