Decompute is the first company in the world to fine-tune powerful AI models locally on as little as 1 GB of VRAM (Checkout our latest blog on AtlasTune). Our flagship platform, BlackBird, lets users build fully on-device AI agents, already downloaded over 20K+ times in its first week—specialized in finance, legal, coding, voice, image generation and beyond. We outperformed cloud-first players like Perplexity and Julius in complex, multimodal, long-form document analysis—while running entirely offline on a Mac Air using our own proprietary models. Our model footprint scales down to just 350M parameters, surpassing Phi and others, and runs efficiently on both CPUs and GPUs.Enterprise edition available now. Contact us for your on-prem/on-device AI needs. Reform, Rise and Reclaim.
External link for Decompute
N Wolfe Rd
Cupertino, California 95014, US
Decompute reposted this
What if AI could learn a new task just by watching? Decompute's Kestrel VLM learns complex visual tasks from only a few demonstrations, bridging perception, reasoning, and action in one model. Built for the edge. Designed for the real world. Thank you to my wonderful team for putting this demo together.
Decompute reposted this
What if AI could learn a new task just by watching? Decompute's Kestrel VLM learns complex visual tasks from only a few demonstrations, bridging perception, reasoning, and action in one model. Built for the edge. Designed for the real world. Thank you to my wonderful team for putting this demo together.
Decompute reposted this
What if AI could learn a new task just by watching? Decompute's Kestrel VLM learns complex visual tasks from only a few demonstrations, bridging perception, reasoning, and action in one model. Built for the edge. Designed for the real world. Thank you to my wonderful team for putting this demo together.
Decompute reposted this
🔥 New Release on high demand: Now train Open AI’s GPT-OSS 20B for personalized agents locally on Mac with BlackBird: ⚡ On-device training 🔒 Full privacy 🛠️ No cloud, no compromise https://lnkd.in/gTpBQStv (We received several messages from everyone requesting this release, and it seems MLX was broken and did not support fine-tuning, so we fixed it ourselves.) For enterprise version, ask us about “BlackBird Enterprise” used by teams from Fortune 500 companies! #BlackBird #Decompute #OpenAI #oss
🔥 New Release on high demand: Now train Open AI’s GPT-OSS 20B for personalized agents locally on Mac with BlackBird: ⚡ On-device training 🔒 Full privacy 🛠️ No cloud, no compromise https://lnkd.in/gTpBQStv (We received several messages from everyone requesting this release, and it seems MLX was broken and did not support fine-tuning, so we fixed it ourselves.) For enterprise version, ask us about “BlackBird Enterprise” used by teams from Fortune 500 companies! #BlackBird #Decompute #OpenAI #oss
Decompute reposted this
Head of Engineering | AI & Cloud Architect | Generative AI & Agentic Systems | FinOps, DevSecOps & MLOps Leader
Looks like AI keeps getting more interesting at the edge. I being testing Blackbird by Decompute it at the most simple an AI that runs locally on your desktop. What caught my attention is how this aligns with NVIDIA’s latest research, where they highlight that Small Language Models (SLMs) can actually outperform LLMs in certain use cases like in terms of speed, efficiency, and adaptability. This may be the start of a big wave, moving from heavy, cloud-only LLMs to smaller, and more personal AI that scales with individual and enterprise needs. Let's see how many companies and professionals will start adopting this SLM-first approach for those curios to read the all papper https://lnkd.in/gHJBFVGM
Decompute reposted this
Head of Engineering | AI & Cloud Architect | Generative AI & Agentic Systems | FinOps, DevSecOps & MLOps Leader
Looks like AI keeps getting more interesting at the edge. I being testing Blackbird by Decompute it at the most simple an AI that runs locally on your desktop. What caught my attention is how this aligns with NVIDIA’s latest research, where they highlight that Small Language Models (SLMs) can actually outperform LLMs in certain use cases like in terms of speed, efficiency, and adaptability. This may be the start of a big wave, moving from heavy, cloud-only LLMs to smaller, and more personal AI that scales with individual and enterprise needs. Let's see how many companies and professionals will start adopting this SLM-first approach for those curios to read the all papper https://lnkd.in/gHJBFVGM
Decompute reposted this
Decompute's Kestrel VLM just scored 841/1000 on OCRB V1, even beating GPT-4o at just 1.5B parameters. Go give it a try at https://www.decompute.run/ PS: We will appreciate your feedback!
Decompute reposted this
Decompute's Kestrel VLM just scored 841/1000 on OCRB V1, even beating GPT-4o at just 1.5B parameters. Go give it a try at https://www.decompute.run/ PS: We will appreciate your feedback!
🚀 Exciting news! We’ve just shipped a new update for the BlackBird SDK, packed with powerful features driven by your feedback. Now, building on-device AI agents is easier and more seamless than ever — all with just a few lines of code. Here’s what’s new: ✨ Support for Retrieval-Augmented Generation (RAG) and fine-tuning on-device across multiple files and folders ✨ Continuous or nightly fine-tuning updates on specified folders for smooth, up-to-date training ✨ Multi-model support — run and combine multiple AI models effortlessly Plus, we’ve fixed some bugs: 🔧 Fine-tuning invocation issues resolved 🎙️ Voice agents and live recording support improvements Try out the limited version on GitHub and start building private, high-performance AI agents on-device today: https://lnkd.in/gGUPyRTH Download our professional version here for Windows and macOS here: https://lnkd.in/gk-aGjH6 Thank you for your feedback and continued support — we’re just getting started! #OnDeviceAI #BlackBirdSDK #AIagents #EdgeComputing #PrivacyFirst