Mirendil Fern 0.1: 10M-token context coding model

Mirendil, an AI research lab, has released Fern 0.1, its first model, with a context window extending to 10 million tokens designed for research and development workflows that process large code repositories and agent execution traces in a single operation.

What changed

Mirendil announced Fern, a language model built from scratch with context capacity up to 10 million tokens. The company states this translates to roughly 23 megabytes of code or approximately half a million lines, enough to hold a codebase the size of Megatron-LM in context simultaneously. Fern 0.1 is the first version, derived from Mirendil-Base after a brief post-training phase.

The model targets six specific R&D tasks: processing larger codebases, agentic coding workflows that execute commands until tests pass, research code including training loops and data pipelines, debugging from extended training logs, reading agent or training traces end to end, and auditing pretraining datasets for duplicates and contamination.

Mirendil built Fern 0.1 on Mirendil-Base, which the company reports was pretrained by a single researcher paired with AI agents over 3.5 months, with total compute spending under $10 million. The pretraining relied exclusively on public sources: GitHub, public code, web pages and public datasets. Mirendil used no distillation, synthetic data, or licensed material. The codebase was written from scratch using PyTorch and JAX.

The current release is available through hosted early access only. Mirendil has announced plans for two larger models: Mirendil-Base-Large between 0.5 and 2 trillion parameters, which is over 80 percent through pretraining, and Mirendil-Base-XL between 2 and 10 trillion parameters, which is over 20 percent through pretraining.

Why it matters

Long-context models remain scarce in the market. While many frontier models advertise large context windows, Fern’s demonstrated ability to effectively use 10 million tokens represents progress for teams working with extremely large codebases or analyzing extended agent execution traces. For R&D teams specifically, holding an entire training run, debugging session, or code repository in context at once eliminates the need for expensive retrieval pipelines or summarization steps that might discard critical information.

On RULER, a standard long-context benchmark, Mirendil-Base achieved 100 percent on single-needle retrieval across all tested lengths from 32K to 10M tokens. Multi-hop reasoning performance declined to 45 percent at 10 million tokens, though the company notes this used only 20 examples per test cell versus 100 elsewhere.

However, the release comes with significant constraints. Availability is restricted to hosted early access, meaning teams cannot deploy it on their own infrastructure or run quantized versions locally. This limits adoption to organizations comfortable with managed access and vendor dependencies.

What to test

Before committing to Fern for production work, teams should verify several claims:

• Long-context retrieval performance in realistic conditions. Mirendil reports 100 percent on RULER single-needle retrieval up to 10M tokens, but performance under real-world noise and multiple queries needs independent validation.

• Agentic coding capability at extended context. The company’s Terminal Bench 4.0 score of 6.1 percent versus 66.4 percent for Claude Opus 5.5 signals a significant gap on intermediate benchmarks. Teams should test Fern on specific coding tasks rather than rely on headline context numbers. On SWE-Bench Verified, Fern scored 73.4 versus 81.6 for Beam.

• Cost per token at 10M context. Early access pricing remains unpublished. Clarify compute and token costs before building workflows around the model.

• Loss improvement generalization beyond test repositories. Mirendil tested next-token prediction on its own repositories and measured cross-entropy falling to 0.155 at a 23-megabyte window. Independent evaluation on diverse codebases would reveal whether this improvement generalizes across different code styles and domains.

• Inference latency at 10M tokens. Long-context processing often incurs latency penalties. Measure end-to-end response time for realistic agent workloads to assess practical feasibility.

The conclusion

Fern 0.1 fills a specialized niche for R&D teams requiring extreme context depth. Its 10-million-token capacity and demonstrated ability to retrieve information across that full range sets it apart from mainstream alternatives. For long-context retrieval and code repository comprehension, benchmark results show promise. For agentic coding tasks, results lag both frontier models and some open alternatives, suggesting Fern is purpose-built for research and analysis rather than general coding work.

The restricted early access model and absence of open weights limit immediate adoption. Teams should request access, test on representative workloads, and confirm pricing before building critical workflows around it. The availability of open weights or broader deployment options would meaningfully expand Fern’s appeal beyond R&D labs willing to use hosted inference.

Related reading

• Kimi K3.1 Pro benchmarks show 3D coding strength
• Agent Skills standard donated to Agentic AI Foundation
• Make vs Zapier: No-Code Automation Platforms Compared