Mirendil, an AI lab founded a year ago, has released Fern 0.1, a model that can hold up to 10 million tokens of context at once. That is roughly ten times the 1-million-token windows of Claude Opus 5 and Gemini 4 Argon, which makes the Fern 10M-token context window the headline of the launch.
The company says that much room fits about 23 MB of code, or around half a million lines. Access is still listed as coming soon, so most people cannot try it yet.
What Mirendil Announced With Fern 0.1
Fern is Mirendil’s first model, and the company says it was pretrained from scratch. Fern 0.1 is that base model after a short round of post-training, and Mirendil says new versions will ship often.
The lab is not pitching a general chatbot. It aims Fern at the work of people building AI, and it lists several jobs the long window is meant to handle.
- Auditing training data for duplicates, benchmark contamination and junk
- Debugging a failed training run by reading long logs and stack traces
- Reading an agent’s full multi-day trace from start to finish
- Working across large codebases and running commands until tests pass
Why a 10M-Token Context Window Matters
A context window is how much text a model can keep in view at one time. Today, when a project is too big, developers often split it into chunks and rely on a retrieval system to find the right piece. A much larger window could make that workaround unnecessary for some jobs.
Mirendil argues that context length is a second way to spend compute, alongside adding parameters. A bigger window lets a model draw on more documents, more code and more of its own past work. If you follow Claude or Gemini 4 Argon, this is the same race, just with a much bigger number.
How the 10M-Token Context Window Performs
Mirendil tested the base model on RULER, a long-context benchmark from Nvidia researchers, at the full 10-million-token length. The company ran 20 examples for each reported result, which is a small sample. All figures below come from Mirendil itself.
| RULER test at 10M tokens | Reported score |
|---|---|
| Single-needle retrieval | 100% |
| Multi-key retrieval | 90% |
| Multi-hop questions | 45% |
The gap between the easy and hard tests is the part to plan around. Finding one fact in a huge pile is far simpler than linking several facts together. At 45%, the model misses more than half of the multi-hop questions at full length.
Mirendil also says more complete long-context results are coming soon. Independent testing will show how well the window holds up in real use.
Who Is Behind Mirendil
Mirendil came out of stealth in June with a reported $200 million seed round at a $1 billion valuation. Andreessen Horowitz and Kleiner Perkins co-led it, with Nvidia also investing. Founders Behnam Neyshabur and Harsh Mehta previously worked at Anthropic and Google.
According to reporting, the company says it built its base model in about 3.5 months with a single researcher and under $10 million in compute. Since that figure comes only from Mirendil, treat it as a claim until others can check it.
What This Means for You
Most readers cannot use Fern today, and no pricing has been shared. Enterprise and research teams are likely first in line, since the focus is AI development work rather than consumer chat.
- Developers: bigger windows may help with huge codebases, but test multi-step reasoning before trusting them. Our guide to an existing codebase shows today’s approach.
- Teams using agents: reading a whole trace in one pass could make debugging simpler, as seen in tools like Claude agent tracing.
- Everyone else: watch whether larger rivals extend their own windows past 1 million tokens.
For more launches like this one, browse our AI news section.
Sources
Primary source: Mirendil’s Fern announcement. Additional reporting from Startup Fortune.
Conclusion
Fern 0.1 puts a 10-million-token window on the table from a well-funded young lab, but the lab’s own benchmark shows weak multi-hop results at that length. Until access opens and outside tests arrive, the number is a promise rather than a proven advantage.
