Mirendil announces a $1 billion fundraising round

Key Takeaways

1Fern can process 10 million tokens per prompt, according to Mirendil

2Mirendil raises $1 billion in Series A, with Kleiner Perkins and Clearlake

3Two versions with up to 10 trillion parameters were presented at the end of last month

Why it matters : Mirendil plans to use these algorithms to develop more efficient LLMs.

The startup founded by former Google DeepMind and Anthropic employees details in-house models used for a coding assistant and claims to have automated a significant portion of the training cycle through agents. It announces a $1 billion Series A round, co-led by Kleiner Perkins and Clearlake, with participation from Google.

In-house Models, a Coding Assistant, and Sizes Up to 10 Trillion

The Mirendil-Base model serves as the foundation for Fern, a coding model unveiled last month. According to the company, Fern is designed to automate programming challenges encountered in AI labs and can handle 10 million tokens per prompt, which, according to Mirendil, is 10 times more than Claude Opus 5.5, equivalent to about 500,000 lines of code. Mirendil states that Fern is based on a version of Mirendil-Base ranging from 100 to 500 billion parameters and that it presented two enhanced versions at the end of last month, peaking at up to 10 trillion parameters. Earlier this year, the company tasked a single researcher with creating a coding model aimed at the level of leading open-source LLMs. Mirendil reports that its automation allowed this project to be completed solo, at a cost of $10 million and over approximately 3.5 months. The company claims that its agents wrote more than 99% of the code for Mirendil-Base, set up the infrastructure, chose the training algorithm, collected the data, and refined both the dataset and the architecture.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Automating Training with Lor and Hundreds of Agents

The first product, Lor, automates many tasks related to training sessions. Mirendil explains that this automation relies on distributing work among hundreds of AI agents. The development of language models proceeds through trials of variants and comparison of results; the company claims that Lor supports this flow and that the platform automatically corrects errors detected during evaluations.

Founders from DeepMind and Anthropic

Founded last year, the company was co-founded by two former researchers from Google DeepMind and Anthropic PBC. Behnam Neyshabur, who serves as CEO, is one of the inventors of the SAM algorithm, developed to optimize model performance on tasks for which they have not been explicitly trained. The CTO, Harsh Mehta, has worked on pre-training projects within both labs, a phase that constitutes the initial step in the AI development flow. Mirendil positions itself as a provider of software to accelerate AI R&D.

A $1 Billion Round and Stated Ambitions

The company announces a $1 billion Series A round, communicated on Thursday. Kleiner Perkins and Clearlake Capital co-led the investment, joined by more than half a dozen players, including Google LLC. Mirendil plans to use the algorithms presented to develop more efficient language models.