Mirendil unveils Fern 0.1 with a 10M-token context window

AI lab Mirendil has introduced Fern 0.1, the first post-trained release of its in-house Mirendil-Base foundation model (0.1T to 0.5T parameters). Rather than scaling static parameter count alone, Mirendil engineered Fern’s architecture, training and serving stack around a 10M-token context window capable of holding roughly 23 MB or 500,000 lines of code at once. That capacity fits a library on the scale of Megatron-LM, dozens of complete agent execution traces or an entire pretraining data shard audited for benchmark contamination and duplicates. Two larger models, Mirendil-Base-Large (0.5T to 2T parameters, over 80% through pretraining) and Mirendil-Base-XL (2T to 10T parameters, over 20% through pretraining), are in development.

Interactive charts on the launch page are rendered in browser canvas and cannot be saved as static images, but they show cross-entropy loss on a 42.9 MB concatenated repository stream dropping steadily from 0.2456 bits per byte at 32k bytes of context to 0.1552 bits per byte at 23 MiB (10.2M tokens), where rival models stop near 1M tokens. On the RULER long-context benchmark at 10M tokens, Mirendil-Base reaches 100% on single-needle retrieval, all-needles multi-key retrieval and variable-tracking recall, 90% on multi-key retrieval and 45% on HotpotQA multi-hop question answering. Mirendil has opened an early-access waitlist and has not announced API pricing.

Key numbers

• Context window: up to 10M tokens (about 23 MB or 500k lines of code) on a 0.1T to 0.5T parameter base
• SWE-Bench Pro v2 Hard & Verified: 76.5% on Pro v2 Hard and 73.4% on Verified (OpenCode harness, 1 attempt)
• SWE Atlas Codebase QnA: 53.2%, ahead of Beam (31.8%) and Nemotron 3.5 (24.2%)
• DeepSWE v1.1 & Terminal-Bench 4.0: 44.2% on DeepSWE v1.1 and 6.1% on Terminal-Bench 4.0