# This 10 MILLION Token AI Is CRAZY... But Is It Real? (Mirendil Fern) **URL:** \ **Category:** Content Creators\ **Tags:** prompt-engineer-48, hefty-llm, machine-learning, coding, tech\ **Created:** [28 October 2026 08:00 UTC](https://www.artofsm.art/t/this-10-million-token-ai-is-crazy-but-is-it-real-mirendil-fern/27024 "2026-10-28T08:00:04Z")\ **Posts on this page:** 2\ **Page:** 1
**Author:** ![merefield](https://www.artofsm.art/user_avatar/www.artofsm.art/merefield/32/74_2.png) [@merefield](https://www.artofsm.art/u/merefield)\ **Post date:** [28 October 2026 08:00 UTC](https://www.artofsm.art/t/this-10-million-token-ai-is-crazy-but-is-it-real-mirendil-fern/27024/1 "2026-10-28T08:00:04Z")
[![](https://img.youtube.com/vi/yz4xSmm1D4M/maxresdefault.jpg "This 10 MILLION Token AI Is CRAZY... But Is It Real? (Mirendil Fern)") ](https://www.youtube.com/watch?v=yz4xSmm1D4M) Mirendil’s new AI model, Fern, claims a groundbreaking 10 million token context window enabling it to process vast codebases or agent activities simultaneously, potentially revolutionizing tasks like debugging and auditing, though these claims remain unverified and access is limited. While Fern shows promising performance improvements with longer contexts and competitive code prediction abilities, independent validation is needed to confirm its practical effectiveness and the true impact of its large context capacity. ---
**Author:** ![merefield](https://www.artofsm.art/user_avatar/www.artofsm.art/merefield/32/74_2.png) [@merefield](https://www.artofsm.art/u/merefield)\ **Post date:** [28 October 2026 08:25 UTC](https://www.artofsm.art/t/this-10-million-token-ai-is-crazy-but-is-it-real-mirendil-fern/27024/3 "2026-10-28T08:25:05Z")
The video discusses Mirendil’s new AI model, Fern, which boasts an unprecedented 10 million token context window—equivalent to about 23 megabytes or half a million lines of code. This capability allows the model to hold and process an entire large codebase or days of agent activity simultaneously, addressing a major limitation in current AI agents that must chunk and retrieve pieces of code due to limited context windows. Fern aims to improve tasks like debugging long logs, auditing training data, and tracing agent behavior by keeping all relevant information in memory at once. Mirendil, founded by a team of 20 researchers from top AI organizations and backed by a $200 million seed round, announced Fern in October after signing a multi-year deal with Google Cloud to use their AI hypercomputer infrastructure. However, many details remain undisclosed, including the exact hardware used, training compute, architecture specifics, and access timelines. The base model, Fern 0.1, reportedly contains between 100 billion and 500 billion parameters and is designed for agentic coding and research tasks, but access is still pending and the claims remain unverified by independent parties. The video highlights Mirendil’s performance charts, which show Fern’s error rates decreasing as context length increases up to 10 million tokens, outperforming other models measured only up to 1 million tokens. Retrieval tests indicate Fern can find single items with 100% accuracy at full context length, though reasoning across multiple facts is weaker. Despite these promising results, the presenter cautions that these are claims from Mirendil and need external validation. The model’s development story is notable: one researcher reportedly built the base model in just three and a half months for under $10 million, using only public data and AI agents to write most of the code and run experiments. Comparisons with other models show Fern’s base version performs competitively but not dominantly on code prediction benchmarks, and it currently lags behind some models on agentic coding tasks. The real breakthrough is the large context window rather than raw performance scores. Mirendil is also working on larger models with up to trillions of parameters, aiming to improve scaling and eventually achieve self-accelerating AI that can partly build future models autonomously. However, these remain future goals rather than current realities. In conclusion, the 10 million token context claim is impressive but unproven as a practical capability until independent researchers can test it on real-world codebases. Fern represents a potential shift in AI model development, emphasizing compute, data, and judgment over large teams and budgets. The video encourages viewers to watch for when access opens, further validation of long-context performance, and reproduction of results outside Mirendil. Until then, Fern stands as a promising but cautiously viewed innovation in AI coding models. ## Useful Links - [Mirendil’s official announcement of Fern](https://mirendil.com/news/introducing-fern/) — Direct source of the main claim about Fern’s 10 million token context window and model specifications. - [Mirendil’s scaling and Google Cloud partnership announcement](https://mirendil.com/news/scaling-self-accelerating-ai-with-google/) — Provides context on the infrastructure and compute environment used for Fern’s training and inference. - [Mirendil company announcement](https://mirendil.com/news/announcing-mirendil/) — Background on the company behind Fern, important for understanding the project’s credibility and scale.