š From GPU ā NPU: on-device AI gets real - hereās what we launched recently ā¬ļø
OmniNeural-4B - worldās first NPU-aware multimodal model (text, image, audio).
š 9Ć faster audio (vs Whisper)
š¼ 3.5Ć faster image (vs SigLIP)
nexaML - run SOTA models locally in 1 line, dead-easyā¦
š Nexa SDK is first to support SOTA VLM, @googleaidevs's Gemma 3n (single-image understanding), with 1-line setup, fully local on any Windows GPU.
So your app can
š° read image-based news & threads
šļø auto-label products in marketplaces
š make screenshots & charts searchable
AI has always been GPU-first. But on-device AI should be NPU-first.
Today weāre launching OmniNeural-4B ā the worldās first NPU-aware multimodal model, natively understanding text, images, and audio.
And introducing nexaML ā a generative AI inference engine that runs models onā¦
Run @OpenAI's gpt-oss model on your Mac with MLX & GGUF, side-by-side
Nexa-sdk now supports GPT-OSS in MLX (Apple-optimized) and GGUF formats, right from your command line with a single line of code.
On an M4 Max, MLX hit ~103 tok/s, about 25% faster than GGUF in our tests.ā¦
š§ Every year I lose hundreds of hours digging through local folders for that one PDF, slide, or doc.
But now thereās Hyperlink ā a fully offline AI agent that instantly searches my files, with citations.
No cloud. No spying. Just my files, my device, my AI.
Public beta ššā¦
š§ Every year I lose hundreds of hours digging through local folders for that one PDF, slide, or doc.
But now thereās Hyperlink ā a fully offline AI agent that instantly searches my files, with citations.
No cloud. No spying. Just my files, my device, my AI.
Public beta ššā¦
š SneakāÆpeek: new NexaāÆSDK Beta
Ship onādevice AI in one command:
⢠Run MLX models locally
⢠Oneāline GGUF quant selection from HF
⢠Multimodal input right in your CLI
This is just the warmāupābigger updates for developers coming soon. š„
š SneakāÆpeek: new NexaāÆSDK Beta
Ship onādevice AI in one command:
⢠Run MLX models locally
⢠Oneāline GGUF quant selection from HF
⢠Multimodal input right in your CLI
This is just the warmāupābigger updates for developers coming soon. š„
We've been quietābut busy building.
Previously, we created Nexa-SDK and on-device AI models like Octopus-v2, OmniVLM, and OmniAudio, sparking excitement in the community. š„
Indeed, on device extends what's possible with AI. We are dedicated to make on-device AI frictionāfreeā¦
OmniVision-968M: a new local VLM for edge devices, fast & small but performant š
it's based on SigLIP-so-400M and Qwen-2.5-0.5B
šØ 9x less image tokens, super efficient
š aligned with SFT and DPO for reducing hallucinations
š„ Apache 2.0 license
Dolphin
discuss: huggingface.co/papers/2408.15ā¦
Long Context as a New Modality for Energy-Efficient On-Device Language Models
This paper presents Dolphin, a novel decoder-decoder architecture for energy-efficient processing of long contexts in language models. Our approach addressesā¦
š Don't miss out on this event! š
šļø Date: Thursday, August 1, 2024
š Time: 8:00 PM - 9:00 PM PST
š Register now: lu.ma/nl91404h
Join us for the "Getting Started with Octopus V2" workshop, featuring our talented engineer Ethan Wang. Whether you're new to theā¦
Our Coral SDK achieves impressive inference speeds, faster than Llama-cpp and MLC-LLM for inference on CPU, and we run the benchmark on Phi3 on @Google Pixel 6. Coral SDK supports @Android, iOS, MacOS, and @Windows with multi-processor and flexible quantization options. Discoverā¦
On-device AI planning agents can be more powerful than previously thought. While virtual assistants like Siri struggle w/complex tasks, MIT & @NEXA4AI from Stanford show that a 3B on-device model can achieve a 97% success rate in planning.
Introducing Octo-planner, a languageā¦
šJoin us for the Super AI Agent Hackathon with @huggingface! š¤
Build innovative AI Agents applications across domains, pushing the boundaries of what AI Agents can do with the latest Octopus models.
š Location: Stanford University/Virtual
š Prizeļ¼$7,000
Register now andā¦
š Introducing our Coral SDK! š
Run LLMs and build AI agents on edge devices with ease.
š¹ CPU & GPU support
š¹ Multiple compression options
š¹ Runs on Raspberry Pi
š¹ Available for Android, iOS, macOS, and Windows
Build powerful AI on the edge! šāØ
Find us:ā¦
š In 2007, the introduction of the iPhone revolutionized the traditional keypad phone.
Now, NEXA AI is taking this evolution of human-computer interaction further, moving from complex UI workflows š„ļø to a more intuitive approach: natural language š£ļø
Just tell our AI agentā¦
šGlad to introduce an interesting project of mine - Kanji-Streamingšan interesting dialogue system powered by StreamDiffusionIO
š«Chatbot will stream out diffusion-generated fake Kanji, instead of original text š
See GitHub repo for all the details:
github.com/AgainstEntropyā¦
182K Followers 63 FollowingBuilding new freedoms of imagination for the world through pioneering research and design. Try Dream Machine for free ā https://t.co/LmWmA4H803
25K Followers 33 FollowingWorld Labs is a spatial intelligence company building Large World Models to perceive, generate, and interact with the 3D world.
908 Followers 333 FollowingCo-founder and CTO at Nexa AI, Industrial Veteran from Google & Amazon, and Stanford alumni. Committed to lifelong learning and advancing AI technology.
364K Followers 8 FollowingVercel provides the developer tools and cloud infrastructure to build, scale, and secure a faster, more personalized web. Creators of @nextjs, @v0, and @aisdk.
1.4M Followers 1K FollowingBuilding @EurekaLabsAI. Previously Director of AI @ Tesla, founding team @ OpenAI, CS231n/PhD @ Stanford. I like to train large deep neural nets.
50K Followers 404 Following@AnthropicAI. Prev. @Google Brain/DeepMind, founding team @OpenAI. Computer scientist; inventor of the VAE, Adam optimizer, and other methods. ML PhD.
163K Followers 0 FollowingInvented principles of meta-learning (1987), GANs (1990), Transformers (1991), very deep learning (1991), etc. Our AI is used many billions of times every day.
54K Followers 0 FollowingWe are building a world class AI R&D company in Tokyo. We want to develop AI solutions for Japanās needs, and democratize AI in Japan. https://t.co/1q07mb3TzE