undefined nº 01
AI news you missed while you were at the beach
Hey,
This is undefined_, issue one: a short letter about running open models in production, written by the Helmcode team. Plain text, a few links, twice a month.
If you just got back from vacation, here’s what AI did while you were away.
// deepseek shipped the story of the month
On August 13, DeepSeek released v4 Pro and, the same day, open-sourced dsh, its own agent harness. The repo is already past 150K GitHub stars.
The short version: an agent runtime where everything is a plugin, including the loop that decides what the agent does next. MIT licensed, model-agnostic, and it runs against any OpenAI-compatible endpoint. Under it sits Cordis, a kernel backed by an 88-page paper with formal proofs from Peking University and DeepSeek.
We wrote a full deep dive: 23 sections, with reading routes from 5 to 40 minutes, from getting it running in 60 seconds to the formal theory, so you can go as deep as you actually need and no further.
// then deepseek raised its prices
Three days after the harness, the bill arrived. Since August 16 DeepSeek bills peak and off-peak, and v4 Pro output went from $0.87 to $3.96 per million tokens at peak, or $1.98 in the cheap window, which is still more than double what it used to cost at any hour. Cache-hit input took the sharper end of the rise, at $0.044 per million on v4 Pro and $0.014 on v4 Flash.
Cache hits are exactly what made v4 Flash 0731 so cheap to run in agent workloads, and OpenCode has reported cache-hit ratios near 96% on its Flash traffic, so the routers are now scrambling to keep serving it anywhere near the old price.
Before you rewrite anything, check your clock: off-peak is exactly half of peak on every tier, and peak runs 01:00 to 04:00 and 06:00 to 10:00 UTC, so if your agents work through the European afternoon they are already on the cheaper half of the day.
Our flat rate at Helmcode didn’t move, and v4 Flash 0731 is now live on Enterprise accounts.
// watermarks are here
Anthropic announced that future Claude models will watermark their text output to comply with the EU AI Act, and since around 190 companies signed the same Code of Practice, expect every big closed lab to follow.
We’ve published a plain-language guide to what that means in practice: where the rule comes from, how it affects you as a company and as a developer, what happens with open-weight models, and what you need in place before December 2, 2026.
Watermarking and the AI Act, explained
// quick hits
- GLM 5.3 is out, post-trained with a heavy cybersecurity focus.
- Qwen 3.8 released open weights for its 27B; the big MoE is still pending.
- Meta shipped Muse Glimmer, a 30B multimodal model for local agents, and Nvidia answered a day later with Nemotron 3.5 Lightning, a 30B MoE with 3B active parameters and a 1M token context.
- xAI launched Grok Bot.
- Nvidia, Meta, Microsoft, OpenAI and Google signed a joint letter arguing that open-weight models should not face broad restrictions. On that one, we agree.
// from our side
NaN, the builder community we run, passed 850B tokens.
We’ve also raised the monthly DeepSeek quotas. On the Helmcode Enterprise plans:
- Starter, 5B tokens a month
- Growth, 15B tokens a month
- Scale, 35B tokens a month
And on nan.builders, from 500M to 2B tokens a month.
Two dates worth saving, both with details to come:
- The first Café Helmcode, September 23 in Madrid
- Inference_ #01, with Hugging Face, October 6 in Madrid
That’s issue one. Let’s see how many new models the next one has to cover.
Borja