On October 27, 2026, startup Mirendil, founded by Anthropic alumni, unveiled its first model, Fern 0.1: trained from scratch with short post-training and a context window of up to 10 million tokens. The company's thesis is formulated as "Context is Intelligence": instead of scaling parameters, entire repositories, training logs, and datasets for deduplication are placed directly into the context. Public figures confirm nearly perfect retrieval on ultra-long contexts, but the model solves agentic tasks significantly worse than the frontier, and access to the early version is available only by application.

image
image

What Happened

On October 27, 2026, Mirendil published the announcement "Introducing Fern — Mirendil's First Model" and opened applications for early access. Fern 0.1 was trained from scratch and underwent short post-training; the context window reaches 10 million tokens, which the company estimates at approximately 23 MB of code or 500,000 lines — equivalent to the Megatron-LM codebase. On the bits per byte metric (cross-entropy of next fragment prediction, calculated on 42.9 MB of Mirendil repositories; lower is better), Fern continues to improve up to a 23 MiB window, i.e., approximately 10.2 million tokens with a value of 0.155, while competitors — Kimi-K3, GLM-5.3, Beam, DeepSeek-V4-Pro-Base, and Nemotron 3 Ultra — are measured up to approximately 1 million tokens, where Kimi-K3 is the most accurate: 0.140 versus 0.165 for Fern. In RULER at 10 million tokens, the model solves retrieval tasks almost perfectly: variable tracking — 100%, multi-key search — 91–95%, but multi-hop QA on HotpotQA degrades from 70% at 32K to 51% at 10M.

Context

Trust in the announcement rests primarily on the team: Behnam Neyshabur was co-head of the Discovery team at Anthropic, and before that led the reasoning direction and the Minerva project at Google; according to the company, his work helped triple the computer-use capability of Claude Sonnet 4.5. Harsh Mehta is named as the second founder in the channel text — detailed public biographical details are not available in the accessible material. The product logic explains the chosen niche: Mirendil promotes the thesis "Context is Intelligence" — training logs, multi-day agent trace records, and entire datasets should enter the context directly, rather than through parameter scaling and external retrieval layers. Fern 0.1 is honestly positioned as a base model after a short training cycle with frequent releases of new versions, and the company itself notes that post-training usually slightly worsens bits per byte for post-trained models like Kimi-K3. It follows that the announcement is primarily a statement of a bet on long context, not a claim to frontier agentic capabilities in the near term.

Why This Matters for the Industry

For the industry, the announcement sets a new product axis — "super-long homogeneous context as an interface": an entire repository, complete training logs, or multi-day agent traces are fed directly to the model, instead of assembling a RAG pipeline. Simultaneously, public figures define the boundaries of the hypothesis: retrieval and tracking on ultra-long contexts are nearly perfect, but agentic execution is significantly worse than the frontier, meaning window length does not automatically convert into code quality and agentic capabilities. Fern 0.1 is not a ready-made component, but a signal: a base model under early access, without API, pricing, SLA, or latency data, and it is too early to incorporate it into a product. If the stated cycle of frequent releases is confirmed, the market will have a chance for independent verification: the main missing data are competitor measurements on windows larger than 1 million tokens and results from pre-trained models of 0.5–2 trillion and 2–10 trillion parameters. On a six-month horizon, if larger versions of Fern maintain the 10-million window, receive full post-training, and an API, the first real "entire repository or entire log in context" pipelines are likely in data preparation, log analysis, and agent trace analysis — with an expected counter-move from competitors.

Why This Matters for Users

The practical interest in the model is currently in the narrow niche of giant homogeneous contexts: reading entire repositories, analyzing training logs, and deduplicating and cleaning datasets. It is too early to use it as a coding agent: on Terminal Bench 4.0, Fern scored 6.1%, while even basic competitors provide twice as much. Access to the early version is arranged by application through the form at mirendil.com/early-access; there are no weights for local execution, nor pricing and SLA. A realistic plan for teams now: submit an application and simultaneously collect benchmark tasks from their domain niche — log analysis, deduplication, codebase navigation — to compare cost and quality with current tools upon gaining access. If the hypothesis is confirmed in subsequent versions, part of the RAG and memory infrastructure for homogeneous data could be replaced by direct long context, but it is not advisable to rely on this in production today.

What Is Still Unknown / Limitations

All benchmarks are self-published by Mirendil without independent verification, and the bits per byte metric is calculated on the company's own set of repositories, leaving room for selecting a convenient methodology. Fern's superiority at 10 million tokens is not a comparison on equal terms: competitors simply do not have published measurements on such windows, and this absence of data is not a confirmed lead. Weights are not disclosed, independent replication is impossible; prices, API, SLA, and latency are unknown, so the model's behavior under real load has not been tested. Open questions until independent measurements remain the agentic gap, the drop in multi-hop QA with increasing window size, and the stated cycle of frequent releases.

Sources

Author

Look at AI, editorial team