New guide: which GPU do you need to serve AI models. FP8, NVFP4, H200, B200 and how much VRAM your model needs. Read the guide →

Dithered illustration: a dense field of noise on the left resolving into four clean bright lines on the right.

// newsletter

undefined

What matters in open AI, filtered by the Helmcode team.

There is more noise than there are hours in the day to keep up with AI. So twice a month we send you a short read on what came out and what actually changes for us. We do not cover everything: only what affects you and what is worth your time.

Controller: Helmcode S.L. Purpose: sending you the newsletter. Recipient: Resend (USA). Detail and your rights in the Privacy Policy.

What lands in your inbox

  • What actually changed in open-weight models, and what it means if you run them yourself. In issue 01: why GLM 5.2 goes into the catalogue and Kimi K3 does not, with the numbers in front of you.
  • Notes from operating sovereign inference in production: costs, failures, and decisions we would take again.

How it reads

From issue 01, on why one model went into the catalogue and another did not:

We ran Kimi K3 and GLM 5.2 side by side, and we may as well say it plainly: K3 comes out slightly ahead on output, and on throughput the two of them end up in roughly the same place. What separates them is compute. […] So GLM 5.2 ships, K3 stays out, and GLM 5.2 becomes the best open model we have ever run on Helmcode.