Mirendil unveils Fern, an AI model claiming a 10 million token context window

Mirendil released Fern 0.1 on October 26, claiming a 10-million-token context window — roughly 23 megabytes or 500,000 lines of code — ten times the 1-million-token windows of Google's Gemini 4 Argon and Anthropic's Claude Opus 5. In company-run RULER evaluations at the full 10-million-token length, based on 20 examples per result, Mirendil reported 100% accuracy on single-needle retrieval, 90% on multi-key retrieval and 45% on multi-hop questions. Mirendil, which 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, says Fern 0.1 is a post-trained version of Mirendil-Base, built in 3.5 months by a single researcher for under $10 million in compute.

A one-year-old AI lab just claimed a context window ten times bigger than Gemini or Claude offer today, while its company-run benchmark reported 45% accuracy on multi-hop questions at the full context length. Mirendil released Fern 0.1 on October 26, and the headline figure is hard to ignore. The model can hold up to 10 million tokens of context, according to the company, which it says is roughly equivalent to 23 megabytes of code or about 500,000 lines, close to the full size of Nvidia's Megatron-LM codebase. For comparison, Google's newer Gemini 4 Argon and Anthropic's Claude Opus 5 both have 1-million-token context windows. If Fern's number holds up under real use, it's not an incremental gain. It's a different category of tool. Context window size determines how much text a model can actually hold in its head at once: an entire codebase, a full legal contract, days of agent logs, without chopping everything into chunks and hoping a retrieval system finds the right piece later. That retrieval workaround, known as RAG, is the industry's current patch for memory limits, and it's exactly what a 10-million-token window promises to make unnecessary for certain jobs. Mirendil isn't a garage startup. It 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 putting in money, according to reporting from The Next Web and SiliconANGLE at the time. Founders Behnam Neyshabur and Harsh Mehta both came out of Anthropic and Google, and the company's stated mission was never general chat. It's building AI meant to help researchers design experiments, tune training runs, and evaluate models faster than a human team can. Fern fits that mission directly rather than chasing ChatGPT's lane. Mirendil says Fern 0.1 is a post-trained version of Mirendil-Base, a foundation model the company says it built in just 3.5 months with a single researcher, for under $10 million in compute. That's a strikingly lean figure for a model of this scale, though the company itself is the only source for it. Mirendil is pitching Fern at auditing training data, debugging failed training runs from logs, and reading an AI agent's full multi-day trace end to end, tasks where a huge, literal window beats clever retrieval because nothing can afford to get missed. ForgeCode claims it beats Claude Code and Codex, but the fine print tells another story https://startupfortune.com/forgecode-claims-it-beats-claude-code-and-codex-but-the-fine-print-tells-another-story/ Tailcall's ForgeCode tops its own TermBench 2.0 leaderboard at 81.8%, but on the independently run SWE-bench Verified it edges Codex 1 by just 0.6 points, 72.7% to 72.1%, a far narrower gap than its marketing suggests. - how ForgeCode compares to Claude Code performance https://startupfortune.com/forgecode-claims-it-beats-claude-code-and-codex-but-the-fine-print-tells-another-story/ - ForgeCode TermBench 2.0 leaderboard score explained https://startupfortune.com/forgecode-claims-it-beats-claude-code-and-codex-but-the-fine-print-tells-another-story/ The benchmark tells a more complicated story Here's the part worth reading past the headline number. Mirendil evaluated the base model with RULER, a long-context benchmark developed by Nvidia researchers, at the full 10-million-token length. Based on 20 examples for each reported result at that length, the company reported 100% accuracy on single-needle retrieval, 90% on multi-key retrieval, and 45% on multi-hop questions. That's not a hidden flaw. It's long-context AI's open secret, and Mirendil published it rather than burying it. Long-context models can perform worse on more demanding retrieval and reasoning tasks than on single-needle tests, and a 10x bigger window on the easy test doesn't mean 10x more usable intelligence on the hard one. A 45% score on multi-hop questions at full length means the model gets those wrong more than half the time. For a lab pitching itself on rigor, publishing that number is a point in its favor. For buyers deciding whether to trust Fern with something like full codebase analysis, it's the number to actually plan around, not the 10 million headline. Access to Fern is listed as "coming soon," with Mirendil taking early access requests as of its launch date and no pricing disclosed yet. Given the company's framing around AI research workflows rather than consumer chat, expect enterprise and research-lab customers, not a public chatbot, to be first in line. Whether Fern finds a real niche or gets swallowed once Gemini or Claude simply extend their own windows past 1 million tokens is the open question, and it's one only actual usage, not a benchmark page, will answer. Also read: A Chinese AI coding tool secretly uploaded users' private code to Alibaba Cloud https://startupfortune.com/a-chinese-ai-coding-tool-secretly-uploaded-users-private-code-to-alibaba-cloud/ • Micron says its AI memory chips are sold out through 2026 and into 2028 https://startupfortune.com/micron-says-its-ai-memory-chips-are-sold-out-through-2026-and-into-2028/ • TSMC posts record quarterly revenue and raises its 2026 forecast again https://startupfortune.com/tsmc-posts-record-quarterly-revenue-and-raises-its-2026-forecast-again/ This article is posted in AI News https://startupfortune.com/category/ai/ , check it out for more related stories. Join the discussion Open in the community → https://startupfortune.com/community/ Almost there. Sign in and your reply posts straight away.