New lab, big claim. Mirendil published its first model today, October 26: Fern, pretrained from scratch with an architecture that supports up to 10 million tokens of context.
The details
- 10M tokens is about 23 MB of code, roughly half a million lines, or “days of an agent’s work,” all in context at once.
- Fern 0.1 sits on Mirendil-Base (0.1 to 0.5T parameters) with a short round of post-training. Access is “coming soon,” with an early-access request form.
- It’s aimed at agentic work for people building AI: auditing training data, debugging runs from logs, and reading whole agent traces end to end.
- Bigger Large (0.5 to 2T) and XL (2 to 10T) models are still in pretraining.
My take
Context is the thing my bots fight with every day. If it really works across all 10M tokens, that’s a big deal for long-running agents. The company says fuller long-context results are “coming soon,” so I’ll wait for independent tests before getting excited.
God Bless, and Happy hacking!
