Fern 0.1

Fern 0.1 is the first product model from the U.S. lab Mirendil (founded by former Anthropic / Google researchers), announced on October 26, 2026. It is Mirendil-Base after a short post-training round. Mirendil claims a context window of up to 10 million tokens (about 23 MB of code, or ~0.5M lines) aimed at people who build AI: auditing pretraining data, debugging training runs from logs, and reading full agent or training traces end to end.

The post Introducing Fern states that Fern is Mirendil’s first model, pretrained from scratch, that Fern 0.1 is the first version with frequent updates planned, and that larger Mirendil-Base-Large (0.5–2T) and Base-XL (2–10T) models are still in pretraining. Access is listed as coming soon with early-access requests open. No public weights, API pricing, or full independent eval write-up appear in the announcement.

Positioning

Fern is framed as a long-context agent model for AI builders—not a consumer chat assistant. Mirendil highlights:

• Agentic coding over larger repositories (run commands, read output, keep going until tests pass)
• Research code such as training loops, data pipelines and evaluations
• Debugging from long logs and stack traces
• Analyzing full agent or training traces end to end
• Auditing whole pretraining shards for duplicates, contamination and junk

Company-reported long-context evidence (read carefully)

• On RULER at 10M tokens for the base model, Mirendil reports ~100% single-needle retrieval and ~45% on multi-hop questions (~20 examples each)—self-reported, tiny samples
• Internal next-token / bits-per-byte curves on Mirendil’s own repos are said to keep improving as context grows; that does not generalize to arbitrary codebases
• The page says more comprehensive long-context results are coming; agentic benchmark scores are largely missing from the announcement text

Relation to prior radar entries

The radar already tracked Mirendil-Base (and skip-listed Base-Large / Base-XL). Fern is the post-trained product name, shipped publicly for the first time, so it gets its own term page.

How to get access today

1. Read the announcement: https://mirendil.com/news/introducing-fern/
2. Request early access on the Mirendil site (still marked coming soon)
3. Watch @mirendil for weights / API; Baseten has been mentioned as an inference partner for base-model evals, but Fern 0.1’s public calling path is not specified in the launch post

Who it’s for

• AI labs and infra teams that need to read huge repos, training logs or full agent traces
• Researchers evaluating the “ultra-long-context coding / research agent” category

Caveats

• No public weights or pricing yet—no local deploy or pay-per-token trial from the announcement alone
• 10M context is an architecture/serving claim; effective use and cost (KV cache, latency) need independent evals
• Size is described as about 0.1–0.5T parameters for the Fern 0.1 / Mirendil-Base tier; larger tiers are still training