On Oct 26 Mirendil unveiled Fern, its first model, pretrained from scratch on its own Mirendil-Base (0.1-0.5T params) with a context window of up to ~10M tokens, about 0.5M lines of code or days of agent traces. Fern 0.1 got only a short round of post-training and targets AI R&D work such as auditing training data, debugging runs from logs and reading agent traces end to end. Vendor-reported: 73.4% on SWE-Bench Verified and 100% single-needle RULER retrieval at 10M, while multi-hop QA drops to 45% at 10M. Access is early-access only; no price or weights yet.
Key Takeaways
- ✓Scale and context (vendor-reported): Mirendil-Base has 0.1-0.5T params; Fern reads up to ~10.2M tokens (about 23 MB or ~0.5M lines of code, roughly Megatron-LM scale), per the official blog
- ✓Coding scores (run by Mirendil, OpenCode, 1 attempt): SWE-Bench Verified 73.4, SWE-Bench Pro v2-Hard 76.5, SWE Atlas QnA 53.2, DeepSWE v1.1 44.2, Terminal-Bench 4.0 just 6.1; Claude Opus 5.5 in the same table scores 98.2/100.0/66.4/74.2/66.4
- ✓RULER long context (vendor-reported): single-needle retrieval is 100% from 32K to 10M; multi-hop HotpotQA falls from 65% at 32K to 45% at 10M, with only 20 samples per 10M cell
- ✓Cost (vendor-reported): one researcher plus agents built Mirendil-Base in 3.5 months for under $10M of compute, with agents writing over 99% of the code, per the pretraining write-up
- ✓Availability: Fern is the smallest of three models; Base-Large (0.5-2T) and Base-XL (2-10T) are still pretraining; access is via early-access requests only, with no price or weights announced
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and industry pain point
Mirendil is a lab founded this year. CEO Behnam Neyshabur previously co-led Anthropic's Discovery team, and CTO Harsh Mehta did pretraining at Anthropic and Google DeepMind. Much of the grind in AI R&D is reading huge training logs, checking data contamination and going through full agent traces, material that often exceeds the ~1M-token windows of mainstream models and has to be chunked. Fern is pitched at putting a whole codebase or days of agent work in context at once.
Architecture and mechanism
Fern 0.1 is the in-house Mirendil-Base (0.1-0.5T params) after a short round of post-training; Mirendil says its architecture, training and serving support up to ~10M tokens. On a 42.9 MB stream of its own repositories, Fern's bits per byte fell from 0.2456 at a 32K window to 0.1552 at ~10M, showing longer context keeps helping prediction. Two larger models, Base-Large and Base-XL, are still pretraining.
Benchmarks (vendor-reported)
All figures are Mirendil's own runs. SWE-Bench Verified 73.4 and SWE-Bench Pro v2-Hard 76.5, level with Reflection Beam but well behind Claude Opus 5.5 and GPT-6 Astra; Terminal-Bench 4.0 is only 6.1. RULER single-needle retrieval stays at 100% up to 10M, while multi-hop QA drops to 45% at 10M on just 20 samples. Mirendil itself says 10M is not yet as usable as 1M.
Getting started for developers
For now there is only an early-access request form; no price, API or weights have been published. It is worth watching for long-log debugging, whole-repo audits and training-data dedup; for everyday coding and terminal tasks, frontier models remain the better choice today.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.