Zack Li
@zacklearner
Followers
327
Following
144
Media
13
Statuses
107
Co-founder and CTO at Nexa AI, Industrial Veteran from Google & Amazon, and Stanford alumni. Committed to lifelong learning and advancing AI technology.
Joined October 2021
We just updated the Nexa SDK for beta launch. Now you can run MLX models, easily pick GGUF quants from Hugging Face, and interact with multimodal inputs—all directly from the CLI. This is just the beginning—bigger updates are on the way.
1
0
4
Thanks @googledevs to feature our OmniAudio-2.6B model!.
🎉 We're thrilled that @googledevs featured our OmniAudio-2.6B model!. ✨ See @Google's full post at:
0
0
2
Thanks @Google @googledevs for promoting our work!.
🎙️📱 Make your apps more conversational and responsive while keeping data private with OmniAudio's on-device audio understanding. Explore the full breakdown → #GoogleDeveloperNews
0
0
1
Catch us at CES! Let’s talk on-device AI solutions for your business:.📍 LVCC North Hall, Booth 9177.📍 Venetian Hotel, Titian Rooms 2303.
🌟 Incredible first day at #CES2025! 🎉 We’re thrilled by the amazing response to our real-time edge AI demos and are beyond grateful for @AMD's support. 🙌🔥 Stop by LVCC North Hall #9177 or Venetian Hall #2302-2305—let’s chat about your on-device AI use cases! 🚀💡
1
0
2
The best is yet to come in 2025! Stay tuned for breakthroughs in model compression, local inference, and on-device multimodal models with Nexa AI. 🚀🔍 Read more:
nexa.ai
Nexa AI's 2024 Milestones and Highlights at a Glance
✨ 2024 was transformative for on-device AI and Nexa AI! From launching groundbreaking products to growing our team, we’ve reached new heights together. 🚀. 🌟 Full report: 🎉 Milestones we’re proud of:.📱 Powering on-device AI across PC, Mobile,
0
0
1
I cannot agree it more, Small language models + on-device edge inference will undoubtedly take off in 2025 @JenniferHli.
It's that time again - Big Ideas for 2025! My take is that we will see small but mighty on-device AI take off next year. Smaller AI models will dominate in volume and usage. Why? 🧵
0
0
4
Introducing NexaQuant.✅ 3X lighter models w/ 100%+ accuracy recovery.✅ 2.5X faster on any device.✅ Multimodal: text, vision, audio.Perfect for running LLMs on edge devices. 🔗 Try it free: 📊 Benchmarks: #EdgeAI #AI2025.
nexa.ai
Works with both text and multimodal models and can be deployed on any devices
Introducing NexaQuant: 🦙 Llama.cpp-Compatible Model Compression with 100%+ Accuracy Recovery. Perfect for bringing Gen AI to edge devices, it features:.✅ 3X lighter models with 100%+ accuracy recovery. ✅ 2.5X speed up on any device. ✅ Multimodal support: text, vision,
0
0
0
Excited to showcase Nexa AI’s on-device multimodal AI at #CES2025 with AMD for NPU acceleration!.🔹 Nexa AI @AMD Booth: Venetian Hotel, Titian Rooms 2302-2305.🔹 Nexa AI Booth: LVCC North Hall, Booth 9177.#OnDeviceAI #MultimodalAI #NexaAI #AMD.
Nexa AI is excited to join the @AMD booth at @CES 2025! Come see the first-ever, on-device multimodal AI, accelerated by AMD NPU, GPU, and CPU. 🔹 Find us here: .📍 AMD Booth: Venetian Hotel, Titian Rooms 2302-2305 .📍 Nexa AI Booth: LVCC North Hall, Booth 9177 . And discover
0
0
1
🚀 Meet OmniAudio: our 2.6B-parameter audio-language model for lightning-fast, on-device inference with Nexa SDK! On Mac M4, it delivers 35 tokens/second, enabling offline voice QA, content creation, and more. The future of multimodal AI is here—built for the edge!#AI #OnDeviceAI.
🔊 Meet OmniAudio-2.6B, the world's fastest and most efficient audio-language model that processes both text and audio inputs. 🏗️ Built upon @GoogleAI's Gemma-2-2b, Whisper turbo, and a custom projector module, its unified architecture enables minimal latency and resource.
0
0
1
With Nexa SDK's Audio Language Model inference engine, you can summarize a 1-minute audio recording in just 3 seconds, Try it out from Hugging Face:.Blog: #AI #Audio #LLM.
nexa.ai
Run Qwen2-Audio on edge devices with Nexa SDK
For the first time, you can run @Alibaba_Qwen's Qwen2-Audio on your local device for Voice Chat & Audio Analysis with Nexa SDK - various quantization options available. 🎉. 👉 🔊 Qwen2-Audio is a SOTA small-scale multimodal model that handles audio and
0
0
1
Excited to collaborate with AMD to further enhance OmniVision, our sub-billion parameter vision-language model, for NPU and iGPU inference. Together, we're pushing the boundaries of efficient, on-device AI solutions!.
Experience OmniVision, our sub-billion-parameter multimodal model, delivering seamless AI capabilities on @AMD-powered devices!. Running on the @AMDRyzen AI 9 HX 370, OmniVision showcases:.⚡ Tiny Model Size: only 968M parameters.⚡ Efficient Processing: Powered by AMD iGPU
0
0
1
🚀 Excited to unveil OmniVision – Nexa AI's 968M Vision-Language Model, benchmarked to outperform SOTA sub-1B models and designed to power visual AI experiences on edge devices. 🔗 Nexa SDK: 🔗 Hugging Face Demo:
huggingface.co
🎉Meet OmniVision, a compact, sub-billion (968M) multimodal model optimized for edge devices. Improved on LLaVA's architecture, it processes both visual and text inputs with high efficiency for Visual Question Answering👀and Image Captioning🖼️:. - 9x Tokens Reduction: Reduces
0
0
2
Happy to share that our on-device AI inference toolkit, Nexa SDK, has achieved 1k star in 2 weeks:.
github.com
Nexa SDK is a comprehensive toolkit for supporting GGUF and MLX model formats. - NexaAI/nexa-sdk
🎉Nexa SDK just hit 1K 🌟on GitHub in 2 weeks! . With Nexa SDK, developers can deploy text 📝, audio 🎵, and vision 👁️ models on-device to build AI applications—no API costs, no internet needed. 🔥 Run llama3.1 (copy & paste into terminal):.pip install nexaai && nexa run
1
0
1
Alibaba just open-sourced Qwen2.5, and we’re thrilled to announce that Nexa SDK now supports Qwen2.5, Qwen2.5-code, and Qwen2.5-Math on the same day! 💡 Dive into the latest features and experience next-level on-device AI with Nexa SDK at :
github.com
Nexa SDK is a comprehensive toolkit for supporting GGUF and MLX model formats. - NexaAI/nexa-sdk
🙌 Nexa SDK now supports Qwen2.5, Qwen2.5-code, Qwen2.5 Math. See changelog for instructions： Qwen2.5 is the latest release in the Qwen large language model series, featuring improved coding, mathematics, and instruction-following
0
0
2
RT @DynamicWebPaige: 🎉Congrats to all of the teams at @agihouse_org's GenAI Goes Local hackathon, and to the talented panel of judges!. 🥇De….
0
9
0
RT @genaisummitsf: 🚀 Excited for GPTDAO's in-person event on September 8th: Unveiling AI Agents in Action with.@nexa4ai ($10M+ Funded) & By….
0
4
0
Nexa AI, in collaboration with Meta and UCSF, has reviewed the latest advancements in on-device LLMs, covering architecture design, model compression, and edge-cloud deployment solutions. paper：.github：
0
0
2
RT @_akhaliq: Dolphin. discuss: Long Context as a New Modality for Energy-Efficient On-Device Language Models. Thi….
0
21
0