IBM watsonx

Behind IBM’s HR AI strategy is something bigger than the tech itself - a culture shift. From shutting down hotlines to piloting agentic AI, we’ve learned that transformation happens when behavior, leadership, and values lead the way. Read the full story from Nickle LaMoreaux: https://ibm.co/6043BJQ1V

🔥 Granite 4 is here, the next generation of our Open Source LLMs. They’re not just “better.” They’re different. Granite 4.0 uses a hybrid Mamba/Transformer architecture that drastically reduces memory use without sacrificing performance. That means strong, enterprise-grade LLMs that run fast; even on lower-cost GPUs. In head-to-head benchmarks, even the smallest Granite 4.0 models beat Granite 3.3 8B. At half the size. And a fraction of the cost. But the story isn’t just about speed and efficiency. Granite 4 is also: - 𝗜𝗦𝗢 𝟰𝟮𝟬𝟬𝟭 𝗰𝗲𝗿𝘁𝗶𝗳𝗶𝗲𝗱: the first open models in the world to meet this new global AI governance standard. - 𝗖𝗿𝘆𝗽𝘁𝗼𝗴𝗿𝗮𝗽𝗵𝗶𝗰𝗮𝗹𝗹𝘆 𝘀𝗶𝗴𝗻𝗲𝗱: so you know the models you’re running are authentic and tamper-free. - 𝗙𝘂𝗹𝗹𝘆 𝗼𝗽𝗲𝗻 𝘀𝗼𝘂𝗿𝗰𝗲 (𝗔𝗽𝗮𝗰𝗵𝗲 𝟮.𝟬): no lock-in, no restrictions. - 𝗘𝗻𝘁𝗲𝗿𝗽𝗿𝗶𝘀𝗲-𝗿𝗲𝗮𝗱𝘆: trusted already by EY, Lockheed Martin, and others for real-world agentic workflows. The models come in sizes as small as 1B active parameters (Tiny) and scale up to 9B active (Small); optimized for use cases from customer service automation to multi-tool agents. And they’re available everywhere: IBM watsonx.ai (of course), Hugging Face, Docker, LM Studio, NVIDIA NIM, Ollama, Replicate, and more. The bigger takeaway? We’re moving past the age of monolithic, one-size-fits-all LLMs. Get Started Here: - Announcement blog:https://lnkd.in/ge5gz9Bc - Hugging Face 3B dense hybrid: https://lnkd.in/exabrV-t - Hugging Face 3B MoE with 1B active: https://lnkd.in/e98GVG_5 - Hugging Face 32B MoE with 9B active: https://lnkd.in/eX4hUdVR - Hugging Face Full Model collection: https://lnkd.in/efwqj939 - In-browser demo: https://lnkd.in/ei_Pyqst

IBM is back! They just joined Hugging Face Enterprise & released Granite 4.0 in open-source with a new hybrid Mamba/transformer architecture that reduces memory requirements without reducing accuracy much. This set of models are great for agentic workflows like tool calling, document analysis, RAG, especially in an enterprise setup 🚀 The "Micro" (3.4B) model can even run 100% locally in your browser on WebGPU, powered by 🤗 TransformersJS! 3B dense hybrid: https://lnkd.in/exabrV-t 3B MoE with 1B active: https://lnkd.in/e98GVG_5 32B MoE with 9B active: https://lnkd.in/eX4hUdVR 🗂️ Full Model collection: https://lnkd.in/efwqj939 🔗 In-browser demo: https://lnkd.in/ei_Pyqst

The paradigm is shifting. Stay tuned to IBM.

📈 Granite-Docling-258M is now the #1 trending model on Hugging Face with 30,422 downloads in ~7 days! Why are so many people paying attention? Because this is a breakthrough in how businesses can finally make sense of their documents. For decades, companies relied on OCR to “read” PDFs, contracts, manuals, or reports. The problem: OCR just pulls out text and loses the structure — tables, equations, diagrams, captions, and layout. That makes it hard to actually use the information in AI systems. Granite-Docling changes that: - It keeps the full structure of a document intact. - It’s compact (258M parameters) yet rivals much larger systems, making it fast and cost-effective. - It outputs data in a machine-readable format, so it can be plugged straight into search, RAG, and enterprise AI workflows. 👉 The bottom line: AI can’t help enterprises if their knowledge is locked inside PDFs. Docling is quickly becoming the open standard for turning documents into usable knowledge and the traction (30K+ downloads in a month) shows how strongly the market agrees. - HuggingFace: https://lnkd.in/gJGNA37e - Video: https://lnkd.in/g9mR_WaR - Blog: https://lnkd.in/gZJuKPub - Tutorial: Build an AI-powered multimodal RAG system with Docling: https://lnkd.in/gaAxRiqm - Tutorial: Build a document-based question answering system by using Docling: https://lnkd.in/g3KYDQ3q - Workshop material: https://lnkd.in/gEEQVKW6