Mirendil Raises $1B Series A for Agent-Led AI Research

The finance hire at Mirendil will have to make sense of a research lab whose agents can keep the machines busy long after a person closes a laptop. The San Francisco company says it is making large compute commitments and beginning to bill enterprise customers. Its new $1B Series A, announced November 12, 2026, gives that accounting assignment considerably more room to grow.

Kleiner Perkins and Clearlake Capital led the financing, with Battery Ventures, Doerr Capital, Gil & Co, Google, M13, Pear VC and S32 among the participants. Mirendil builds models and software for AI research and development, including its private-beta agent environment Lor. The commercial proposition is ambitious: make work associated with a large AI laboratory available to smaller research and engineering teams, while giving those teams more experiments and machine time to direct.

What Mirendil's Series A funds

The funding follows a $200M seed announced in June, co-led by Andreessen Horowitz and Kleiner Perkins, with NVIDIA subsequently named as an investor. Mirendil has not disclosed a Series A valuation in its latest announcement. The release connects the new capital to expanding access to frontier AI R&D, without supplying a spending allocation or a schedule for general availability.

Four co-founders sit behind the effort: CEO Behnam Neyshabur, CTO Harsh Mehta, Shayan Salehian and Tara Rezaei. Kleiner Perkins' account of the team describes Neyshabur and Mehta working together at Google and later Anthropic, where automated research became part of their work. Salehian brought experience from xAI and Twitter, while Rezaei had researched at OpenAI. That history helps explain why the company is trying to package the surrounding laboratory as well as a model.

For a prospective customer, the purchase would reach into how research gets organized. A lab needs to choose training data, launch jobs, examine failures and decide which results justify continuing. Mirendil's opportunity is to make more of those activities available through a connected product, so a customer can spend less effort assembling the machinery around each experiment.

How Mirendil-Base, Fern and Lor fit together

Mirendil's public work separates into a pretrained model, a model adapted for agent tasks and the environment where agents operate. In its Mirendil-Base research account, the company says one researcher built its initial coding model in 3.5 months using less than $10M in compute. Agents performed much of the engineering, while the researcher selected directions, evaluated results and intervened when needed. The reported result concerns a specific internal project with an experienced researcher.

Fern adds post-training and supports a context window of up to 10M tokens, according to Mirendil. A large context window gives an agent room to carry more code, documents and prior work into its next step. Mirendil's published page also records corrections to two training-result figures, a useful reason to read the current technical account rather than circulate an older chart.

Lor is the working environment: agents write code, conduct experiments and coordinate tasks across machines. Mirendil describes runs lasting days and the ability to allocate compute as work requires it. Lor remains in private beta, so its published capabilities should be read alongside that access limit. The combination could appeal to a research group that needs both model capability and the practical means of running repeated experiments.

Compute becomes part of the buying decision

Google participates in the financing while Google Cloud supplies infrastructure for Mirendil's research. The Google Cloud partnership covers training and research infrastructure using TPUs and NVIDIA systems. A supplier of the machines and an investor in the company can therefore have a stake in the same expansion, even though neither relationship discloses what an individual customer's experiment will cost.

Mirendil's current compute vacancy makes the daily work legible. The role includes fleet reliability, scheduling, capacity decisions and understanding what each GPU-hour produces. More agents create more opportunities to use a fleet, and more demands on the people deciding which workload should receive capacity. An experiment that is easy to launch can still consume a scarce resource.

That operating question connects with the infrastructure demand described in DevCuration's coverage of Mount Thor's Apple-based AI systems. Different research tasks need different environments, and those environments have to remain usable while the software changes. For Mirendil, the product and the supply of compute will have to develop together as outside teams bring their own workloads into Lor.

Selling a research workflow takes a pilot

A founding GTM vacancy asks a technical hire to identify customer problems, design evaluations, agree on success criteria and convert pilots into paid adoption. It also assigns pricing and procurement work. Those are concrete steps between an impressive internal experiment and a buyer willing to keep paying for the workflow.

The opening is especially relevant because research groups can disagree about what counts as progress. A faster experiment matters only in relation to the question being investigated and the confidence the team needs in its answer. Mirendil's data vacancy explicitly leaves the training mixture and decisions about trustworthy results with a human, while agents run pipelines and tests. The customer's expertise remains part of the work the software has to support.

The finance vacancy adds another view: Mirendil says enterprise billing is beginning as compute commitments expand. It provides no revenue or named-customer figures. As Lor reaches more teams, researchers, commercial staff and finance will be working through the same practical question from different sides: which experiments help a customer's science enough to keep the machines running for them?