Community Forums

We’re doing exploratory research on how AI teams actually run scaled workloads. Want to contribute? Our team at HP is running a short, self‑guided AI‑moderated interview to better understand how teams approach AI development, model training, and production inference—and where today’s compute setups fall short. We’re especially interested in AI/ML engineers, data scientists, MLOps or platform engineers working with GPUs. If your use cases align, you can also opt in to be considered for future lighthouse collaboration. What to expect:  ✅ ~15–20 minutes ✅ No scheduling — complete it anytime  Take the interview here: 👉 https://tinyurl.com/hpincai 

Mistral AI unveiled Voxtral Mini Transcribe V2 and Voxtral Realtime, speech‑to‑text models aimed at enabling near real‑time multilingual transcription and translation with low latency (~200 ms) and support for 13 languages.At ~4 billion parameters, these models are compact enough to run locally on devices like phones and laptops, enhancing privacy and lowering reliance on cloud services. Voxtral Realtime is open‑source under an Apache 2.0 license and designed for live use cases, while Mini Transcribe V2 excels at batch jobs with diarization and timestamps. Mistral positions these tools as cost‑effective, optimized alternatives to larger proprietary systems from major US tech firms, focusing on efficient model design and dataset quality rather than raw scale.The models align with Mistral’s broader strategy to offer regulation‑friendly AI solutions that fill niche and specialist roles in the speech AI space.Top 3 benefits for an AI developer:• Low‑latency, real‑time multilingual transcri

Today HP released the ZGX Toolkit, a VS Code Extension that streamlines AI development workflows on HP ZGX desktop AI supercomputers by providing automated setup of essential open-source AI tools and seamless device discovery on your LAN. Discover, set-up, and start using a ZGX as an AI companion to unlock the compute your workflows need.With the ZGX Toolkit now you can:discover and connect ZGX devices on your network; no manual IP / SSH set-up quickly install popular open source AI packages and apps, including dependencies leverage templates to quickly test prototype and fine-tuning on your ZGX deviceThe toolkit is free and can be installed directly from the VS Code Marketplace. Check out this 1 min video to see how.For more information, visit www.hp.com/zgx-onbard. 

I built a maritime surveillance AI for an upcoming HP ZGX Nano demo with the U.S. Navy. It processes reconnaissance imagery in under 5 seconds—fully local. No cloud, no network, no data leaves the device.Federal agencies face a real gap: commercial laptops can’t run strong vision-language models locally, but classified imagery can’t go to cloud APIs. That leaves analysts with either no AI or unusable AI. This demo closes that gap.Upload an aerial image to the demo and the system identifies vessel type, assesses cargo and activity, then produces a structured threat assessment with actionable recommendations.It uses Salesforce’s open-source BLIP-2 VLM for visual understanding and TinyLlama to generate natural-language intelligence reports. Everything runs on an HP ZGX Nano—smaller than a desktop PC, with 128GB unified memory—deployable on ships, mobile command centers, or in disconnected SCIFs.Models are open source. No classified data—only public imagery and synthetic coordinates. Code

I followed Curtis Burkhalter, Ph.D.’s HP ZGX Toolkit Microsoft VS Code extension playbook posts to link an HP ZGX Nano with NVIDIA Blackwell GB10 GPU to an HP Omnibook 5 with a Qualcomm SnapDragon - instantly unlocking the GPU power I needed.- Paired my laptop with a ZGX Nano companion device- Went from 45 TOPS GPU to a 1,000 TOPS GPU- Installed ComfyUI to use a Hunyuan image → 3D model for inference- Auto-generated STL files- Sent straight to an HP 3D PrinterNo cloud. Just local Edge AI. Connected and seamless.See the 3 min video on LinkedIn  

Every IT leader knows the tension: your teams want to move fast on AI. Your security team needs to vet everything. Those two timelines rarely align.We designed the HP ZGX Toolkit VS Code extension as an open source tool to close that gap.✅SSH key-only authentication — no passwords. The extension handles key generation and configuration automatically when setting up the ZGX as a companion compute device.✅ Digitally signed packages via HP Software Security (HPSS), providing verifiable authenticity for compliance reviews.✅ Open-source code available on GitHub for security audits before deployment.The goal: accelerate procurement and security reviews, not slow them down.This is post 3 of 3 on the ZGX Toolkit release and our partnership in building the HP ZGX Nano with NVIDIA.Explore for yourself!HP ZGX Nano: www.hp.com/zgx and HP ZGX Toolkit: www.hp.com/zgx-onboard

HP might not be the first name you think of in AI. We're here to change that and we're starting by building in the open.The ZGX Nano is built on NVIDIA's reference design — the same foundation Dell Technologies, ASUS, Lenovo, and other OEMs use for their AI workstations. We're not running from that. We're leaning into it.We open-sourced the ZGX toolkit so it's not just for HP hardware. Inspect it. Modify it. Extend it. Run it on whatever hardware fits your environment.The curated tools we automate — MLflow, Podman, Gradio, Streamlit, modern Python tooling — are all open source. Now the automation layer is too.This is how we think AI tooling should work: open, extensible, and not locked to a single vendor.Check out the GitHub repo: https://github.com/HPInc/ZGX-Toolkit This is post 2 of 3 on the ZGX Toolkit release.

Every AI developer knows the feeling: new machine, high-end GPUs, endless setup.You're not training models on day one. You're resolving dependency conflicts. Configuring experiment tracking. Getting your Python tooling sorted. Troubleshooting why three tools that should work together... don't.The hardware was ready. The environment wasn't.Here at HP we heard this over and over in 40+ customer discovery calls as we geared up to build the new HP ZGX Nano in partnership with NVIDIA. Rather than let this opportunity pass us by, we did something about it.With the latest release of the HP ZGX Toolkit, we automate the full developer workflow setup. One-click installation: MLflow, Podman, OpenWebUI, modern Python tooling (uv, poetry, Miniforge), system monitoring, and more. Automatic dependency resolution. Minimal configuration conflicts.From unboxing to running your first model: approximately 30 minutes.Learn more at www.hp.com/zgx-onboard This is post 1 of 3 on the ZGX Toolkit release.

Vincent Caldeira, CTO APAC at Red Hat, experts in making open source accessible, shared an initial review of HP’s ZGX Nano on LinkedIn. The cliff note is that …HP’s ZGX Nano delivers datacenter-grade silicon for local AI development with unified 128GB memory, seamless VS Code onboarding with the ZGX Toolkit, and impressive 20B-model performance—~30 tok/s, high GPU utilization, stable latency—enabling fast, cloud-free iteration for serious AI developers.Check out the full article on LinkedIn 

In September, NVIDIA published an article outlining the value of custom SLMs for Agentic AI working on-prem on machines like HP Z workstations.Key Takeaways:1. Economics of Small Agentic Models (SLMs):Small Language Models (under ~10B parameters) can replace generalist LLMs for most agentic tasks. They’re 10–30× cheaper to serve, require fewer GPUs, and can be fine-tuned in hours instead of weeks. This shift lowers operational costs, reduces latency, and improves sustainability—critical as AI agents scale across industries.2. Localization and Democratization:SLMs enable local, domain-specific adaptation—they can be easily fine-tuned to meet regional compliance, cultural norms, or specialized workflows. This encourages on-prem and edge deployment, supporting sovereign data control and fostering diversity by allowing more organizations to build their own tailored agents.3. On-Prem Model Creation and Management:With frameworks like NVIDIA Dynamo and ChatRTX, real-time, offline inference b

In January I had the pleasure of working with Anbu Valluvan, an HP AI Studio tester, to resolve a GPU limitation issue his team was running into on their personal laptops when Google Colab GPU was not available due to budget constraints or GPU queuing. The team had collected, created data strategies to ingest up and process blood pressure datasets in appropriate sizes to train small models for use with edge devices. However, their access to GPU was limited. To achieve the accuracy requirements for the paper to qualify for peer review and acceptance, Anbu reached out to collaborate on the project using AI Studio because I had access to a NVIDIA GeForce RTX 4070 GPU which allowed him to experiment and prepare the project with multiple iterations when access to NVIDIA A100 GPU was limited.In a couple of minutes Anbu invited me to his AI Studio project, I synced assets and artifacts, and trained the model on my GPU. Because AI Studio saved artifacts, results, and cloned code to Github, Anb

OpenAI released GPT-4o a conversational chat-bot last week.It is integrated into the online chat-bot as well as an API for integration with 3rd party systems. Early findings illustrate it is twice as fast and half the cost, with higher rate limits for better performance. https://vimeo.com/945586717 The multi-group query decoder architecture performs well based on zero-shot prompts, which are plain prompts that do not provide context in the prompt to help the LLM improve its response.  The API leverages the existing OpenAI chat completion API, which requires an API key to load and interact with the model. Many additional parameters, such as temperature, can be tuned.  from openai import OpenAIOPENAI_API_KEY = "<your-api-key>"client = OpenAI(api_key=OPENAI_API_KEY)response = client.chat.completions.create(  model="gpt-4o-2024-05-13",  messages=[    {"role": "system", "content": "You are a helpful assistant."},    {"role": "user", "content": "Who won the world series in 2020?"},