What happened
Mirendil’s announcement, published November 12, 2026, says the company raised $1 billion in a Series A. Kleiner Perkins and Clearlake Capital led the round. Mirendil names Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC, and S32 among the participating new and existing investors. The announcement does not give a valuation, the ownership terms, or a breakdown of how the capital will be spent. It describes the purpose as building tools to make frontier AI research and development more accessible to scientists and engineers.
The company presents the raise alongside a set of results it says it shared over the preceding three weeks. Its central example is that one researcher, working with agents, pretrained Mirendil-Base, described as a competitive coding model, from scratch in 3.5 months, using less than $10 million in compute. Mirendil says it then post-trained that model into Fern, which supports up to 10 million tokens of context. The company characterizes that context length as sufficient for days of an agent’s work. It does not specify in this announcement the exact training recipe, compute accounting, evaluation set, or the comparison behind the word “competitive.”
Mirendil also describes Lor as the product intended to bring this way of working to other users. Lor is in private beta, according to the announcement. The company says it enables hundreds of agents to write code and run experiments for days at a time, and says Mirendil itself uses Lor every day. The announcement does not describe beta eligibility, pricing, model access, user limits, or what safeguards and review steps govern these long-running tasks.
The company frames the round as a way to continue building toward a future in which one person can undertake work that previously required a team, and teams can take on work that once required institutions. It invites prospective employees and people interested in building with Mirendil to contact the company. This is a financing and company-direction announcement; it is not a general availability notice for Lor or Fern.
What the archive adds
New Runtime covered Mirendil’s private-beta launch of Lor on October 31, 2026, describing a workspace that runs Fern alongside other frontier and open-source models and supplies agents with on-demand compute. The Series A is a new event, while the claims about Fern and Lor extend a known story: the company is now tying its research-and-product thesis to a major financing round. The announcement does not establish that the beta has expanded or that access terms have changed.
Trend signal
This does not yet fit a pattern New Runtime tracks. The announcement argues that agent labor can change the scale of model development, but it does not provide independently checkable task-completion costs, evaluations, or operational details that would show a shift in the economics of verified work. Its training-time and compute figures are useful claims to test against future disclosures, rather than sufficient evidence for a broader pattern.
Why now
Mirendil says it has shared its first results over the three weeks preceding the announcement and presents this financing as a way to extend that work. The timing therefore joins a new funding event to the company’s recent claims about model training and its private-beta workspace. The announcement does not set a timeline for spending, product release, or a next research milestone.
Why it matters
A $1 billion Series A can give Mirendil substantial room to pursue a demanding research and product agenda. More consequential for builders is the proposition the company is making: agents might let a small group carry out a longer sequence of coding and experimental work than its headcount alone would allow. If that proposition holds up, teams may need to think about the design of research workflows, experiment review, and the infrastructure that keeps agents useful over extended runs—not simply about adding model calls.
For now, the announcement offers a company-reported example, not a reusable operating recipe. It does not explain how the researcher selected tasks, verified generated code, handled failed experiments, or controlled compute. Lor’s claimed ability to run hundreds of agents for days could become relevant to practitioners if the beta exposes those mechanics and provides evidence of reliable outcomes. Until then, the raise signals the resources and ambition behind Mirendil’s approach, while the practical value remains unproven outside the company’s account.
Limits and uncertainty
The funding amount and investor list are stated by Mirendil; the announcement supplies no valuation or deal terms. The training duration, compute spend, model competitiveness, context capacity, and Lor’s operating scale are also company claims, and the announcement does not provide independent evaluations or reproducible artifacts. It does not disclose how “compute” is calculated, what quality bar Mirendil-Base met, or how many experiments succeeded. Lor remains in private beta, with no public pricing, access criteria, or release schedule stated here. The claims may be accurate, but their comparability and repeatability cannot be assessed from this announcement alone.
Reader takeaway
Treat this as a major financing event paired with a consequential but company-reported claim about agent-assisted model development. The next useful evidence is not another scale claim: it is reproducible evaluation, transparent accounting of compute and human work, and a clearer account of how Lor controls and verifies long-running experiments.