← Back to live feed1 story
Mirendil has launched Fern 0.1, its first AI coding model, which supports up to 10 million tokens and can process roughly 500,000 lines of code in a single context window. The model is a post-trained version of Mirendil-Base, a foundation model built in 3.5 months with one researcher and less than $10M in compute, and is designed for high-volume tasks such as auditing training data and reading agent traces end to end.
Sign in to suggest
Key sources
- SOURCE@bneyshabur“That’s ~0.5M lines of code (roughly Megatron-LM scale) or days of an agent’s work, all in context at once.”x.com
- SUPPORT@bneyshabur“Fern is the first and smallest of three. Two larger ones are being pretrained now.”x.com
- SUPPORT@bneyshabur“Fern’s architecture, training and serving support 10M tokens (~23 MB of code) at once, and it uses them: prediction keeps improving up to 10M.”x.com
- SOURCE@bneyshabur“Fern is built for the high-volume agentic work of anyone building AI: auditing training data, debugging training runs from logs, reading agent traces end to end.”x.com
- SOURCE@baseten“Mirendil-Base-Large had the lowest bits per byte (a tokenizer-independent measure of perplexity) of any model we tested.”x.com
- SUPPORT@bneyshabur“Our first pretrained model was built by one researcher in 3.5 months, with less than $10M in compute.”x.com
- SUPPORT@roberttlange“Agents wrote over 99% of the code and ran the experiments, while the researcher chose directions and decided which results to trust.”x.com
- SUPPORT@ziv_ravid“pre-training coding models with 1 researcher, less than 10M$ compute, kimi k3 level model (with 2T parmas).”x.com