33 articles
Spain's data protection authority has warned a company that wanted to score CVs with AI. The case shows something many teams miss: when an application built on someone else's model makes you a provider under the AI Act.
DeepSeek wants us to stop watching every session. I went through its prototype and its experiments to understand how it plans to do that.
A Claude Code session is about 174 requests of 100,000 tokens each. What that does to a GPU, why with agents compute runs out before memory, and the same maths on a 16 GB card.
DeepSeek open-sourced DeepInfer, the engine it serves V4.1 Flash with. How it works, what it optimises and how to read its benchmarks.
Open models now compete with closed ones on quality at a fraction of the cost. Why many companies still hold back, and what changes when you separate the model from who runs it.
DeepSeek updates V4.1 Flash: up from 39 to 45 on the Artificial Analysis index, with cost per task down 54%. We put it to work on a website and a video.
Iris-3B, an open 3B image model that generates in pixel space with no VAE, against GPT Image, FLUX and friends on the BudgetPixel Arena prompts.
If you use Claude, GPT or Gemini today, this is the open model we would test first for the same job, compared by model class and by workload.
The prompt you type is only a fraction of what an AI system sends to the model. What actually travels, where it crosses, and what stays on the other side.
Knowing where your intelligence runs, what happens to your data and whether you can still change those decisions. A framework for AI sovereignty.
Live on Helmcode, replacing V4 Flash at no extra cost. It redesigns inference for a workload that keeps growing: agents that read far more than they write.
Record and transcribe your meetings with open models, turn them into structured data, and build a searchable memory of every conversation you have had.
GLM 5.3 ties Kimi K3 at the top of the open models and runs on our European infrastructure with 1M of context and the same 150 euros a month.
An 8x H200 node serving GLM-5.2 costs €25,250 a month before it serves a single token. The full numbers, the break-even, and when owning the hardware wins.
Eleven open models solve the 80 enterprise use cases in our guide. The three axes that decide which one, and fifteen cases with theirs next to them.
The five theorems in the Cordis paper, the plugin framework behind DeepSeek Harness, explained in plain language and with what already runs in production.
What a small town hall actually automates with open models: filings, minutes, notices, grant calls and by-laws, with human review kept in place.
Anthropic is going to watermark Claude's text, and the reason is Article 50 of the AI Act. What it requires, who it binds, and what changes with open models.
How to design multi-model systems where one open LLM plans, another executes, and the context they hand each other stays efficient.
GLM 5.2, the 744B MIT open model, now runs on Helmcode's European infrastructure with zero logs and a flat rate of 150 euros a month.
The calm version of the AI Act: which obligations really apply to you from 2 August, which ones the Omnibus moved, and what to have ready.
Upload your documentation once or connect a Git repository, then ask it from the chat. Every answer cites the document it came from.
FP8, NVFP4, H200, B200: what each NVIDIA generation accelerates in hardware, how much VRAM your model needs, and which GPU to buy. With real production data.
The difference between an image that screams 'AI' and a professional one isn't the model, it's the prompt. The complete guide to generating great images in Helmcode.
New hire at Helmcode: communication, marketing and community. Who I am, what I'm here to do, and what we're planning for the NaN community.
How we moved Qwen3.6-35B to NVFP4 on a single RTX PRO 6000, why it kept crashing, and the one-line fix that turned out to be a cuDNN bug, not a vLLM bug.
117 billion tokens, 3.68 million requests, 21 countries, and 99.98% uptime. NaN is a community of builders with its own inference infrastructure and a private platform to deploy apps and agents.
In this post we'll learn what parameters and quantization are, so we can figure out how much space AI models take up.
In this post I'll walk you through how the community's inference servers are set up: the hardware we use, the stack we run, and the models we serve.
I've spent several hours over several days documenting and optimizing my entire local environment so I can "mechanize" the work I do every day managing infrastructure for multiple startups.
This post isn't meant to be a guide on how to use Clawd, but rather a look at how we're rolling it out at Helmcode to have an AI Agent that helps us with our day-to-day work managing the Cloud infrastructure of multiple startups.
Kubernetes is one of the most widely used infrastructure tools among companies, and it has become the standard for running containerized applications at scale all over the world.
Before we start, a bit of context. The infrastructure is hosted on AWS and the architecture was based on Serverless services:
// cookies
We only use strictly necessary cookies to run the site. No analytics, no advertising, ever — see our Cookie Policy.
// preferences