Mirendil Fern: A 10M-Token Context Model for AI Labs

Mirendil Fern is the new long-context flex in AI, and the number is huge: up to 10 million tokens. Mirendil, a lab that came out of stealth this year, introduced Fern on Monday, October 26, calling it its first model, pretrained from scratch. The first version, Fern 0.1, is aimed at high-volume agent work, especially the messy job of reading logs, traces and big codebases. Access is listed as "coming soon," with early access requests open.

What actually happened

Mirendil's launch post lays out a family of models it calls "the grove." Fern 0.1 is the efficient one. A bigger model on Mirendil-Base-Large, listed at 0.5 to 2 trillion parameters, is "in pretraining" at over 80%. The top model on Mirendil-Base-XL, listed at 2 to 10 trillion parameters, is over 20% through pretraining. Neither bigger model has a release date.

The company frames its thesis as "context is intelligence." Its argument: models get smarter with more parameters or more context, and a bigger window lets an agent use more of its own past work, more instructions and more evidence at once. Mirendil says Fern's architecture, training and serving all support the full 10M window.

The use cases are very builder-focused. Mirendil lists agentic coding that runs commands until tests pass, debugging failed training runs from logs, reading full agent traces end to end and auditing pretraining data for duplicates, benchmark contamination and junk.

Who is behind Mirendil Fern

Startup Fortune reported that Mirendil launched out of stealth in June with a $200 million seed round at a $1 billion valuation, co-led by Andreessen Horowitz and Kleiner Perkins with Nvidia also investing, citing earlier reporting by The Next Web and SiliconANGLE. Founders Behnam Neyshabur and Harsh Mehta came from Anthropic and Google.

One claim stands out. Per Startup Fortune, Mirendil says Mirendil-Base was built in 3.5 months by a single researcher for under $10 million in compute. That would be very lean for a model this size, and the company is the only source for it.

The benchmark fine print

The 10M headline needs context. Startup Fortune noted that Google's Gemini 4 Argon and Anthropic's Claude Opus 5 both offer 1 million tokens, so Fern claims ten times more. But the multi-hop score at full length is 45%, meaning the model gets those harder questions wrong more often than right.

Single-needle tests, where the model finds one planted fact, are the easiest long-context check. Multi-hop tests, where it must connect several facts, are closer to real work. Mirendil published the weak number instead of hiding it, which is a good sign, and it says more complete long-context results are coming.

Twenty examples per task is also a small sample. Treat these numbers as early signals, not a final grade.

Cost is the other open question. Holding millions of tokens in context is expensive to serve, and Mirendil has not shared pricing or speed numbers. Until those land, nobody can say whether Mirendil Fern is practical for daily work or only for special jobs where missing a detail is not an option.

A 10M window is a bigger desk. Whether the model can actually find and connect the right papers on it is the real test.

Where it fits

Long context is turning into a race. We covered Google's flagship in Gemini 4 Argon, and open-weight rivals like Kimi K3.1 Pro keep pushing coding agents forward. Mirendil Fern is a different bet: not a chat app, but a tool for AI teams who drown in logs and traces.

What to do as a reader

If you run training jobs or agent pipelines, request early access on the Mirendil site and test Fern on your own long logs. Measure multi-step questions, not just "find the needle."

Do not expect it soon in consumer apps. There is no public pricing and no general release date yet. Also ignore any "Fern" or "Mirendil" token on a DEX. The lab has not announced one, and anything trading under that name is unrelated.

Sources: Mirendil, Startup Fortune and explainx.ai.