Mirendil

A frontier lab building systems that excel at AI R&D, democratizing access to advanced AI research.

Cover Block

Public sources

Field	Value
Name	Mirendil
Tagline	A frontier lab building systems that excel at AI R&D, democratizing access to advanced AI research [Mirendil]
Headquarters	San Francisco, United States [Mirendil]
Founded	2025 [WebProNews, November 2026]
Stage	Series A [Yahoo Finance / PR Newswire, November 2026]
Business Model	SaaS
Industry	Deeptech
Technology	AI / Machine Learning
Geography	North America
Growth Profile	Venture Scale
Founding Team	Co-Founders (4): Behnam Neyshabur, Harsh Mehta, Shayan Salehian, Tara Rezaei [TechCrunch, August 2026] [LinkedIn, Retrieved 2026]
Funding Label	$100M+
Total Disclosed	~$1,200,000,000, comprising a $200,000,000 seed round in June 2026 and a $1,000,000,000 Series A in November 2026 [WebProNews, November 2026] [Yahoo Finance / PR Newswire, November 2026]

Executive Summary

PUBLIC Mirendil is building software to automate AI research and development, and it merits investor attention now because it has paired that thesis with unusually large early financing and a high-profile infrastructure partnership within roughly a year of founding [Yahoo Finance / PR Newswire, November 2026] [TechCrunch, August 2026] [WebProNews, November 2026]. Public reporting places the company in San Francisco and dates its start to late 2025, with Behnam Neyshabur and Harsh Mehta leaving Anthropic before launching the business alongside Shayan Salehian and Tara Rezaei [WebProNews, November 2026] [LinkedIn, Retrieved 2026] [Business Insider, February 2026].

The product story is centered on Lor, a private-beta platform that the company says coordinates hundreds of agents to write code, run experiments, and provision compute over extended periods, with adjacent model work including Mirendil-Base and Fern [Yahoo Finance / PR Newswire, November 2026] [Mirendil, Retrieved 2026]. The differentiation, according to the company and surrounding coverage, is not a general-purpose chatbot but an integrated AI R&D workflow aimed at compressing the labor and time required for frontier model development [Yahoo Finance / PR Newswire, November 2026] [SiliconANGLE, November 2026].

The founding bench is the clearest asset in the public record. Neyshabur is publicly tied to prior senior research work at Google DeepMind and Anthropic, Mehta to DeepMind and Anthropic pretraining and research platform work, while Salehian and Rezaei are linked in public sources to xAI and OpenAI-adjacent experience respectively [LinkedIn, Retrieved 2026] [Behnam Neyshabur CV, Retrieved 2026] [The Videshi, June 2026] [NBC News, Retrieved 2026] [tarark.com, Retrieved 2026].

Funding has moved at a pace that materially changes the company's profile. Mirendil announced a $200 million seed round in June 2026 led by Andreessen Horowitz and Kleiner Perkins, then a $1 billion Series A on November 12, 2026 led by Kleiner Perkins and Clearlake Capital, with Google, Nvidia, Battery Ventures, M13, Pear VC, S32, Doerr Capital, and Gil & Co among disclosed backers; the business model is categorized here as SaaS, though commercial details remain thin in public sources [WebProNews, November 2026] [Yahoo Finance / PR Newswire, November 2026].

The next 12 to 18 months will turn less on fundraising capacity and more on proof of product adoption, repeatable deployment, and whether the Google Cloud relationship converts infrastructure access into customer traction rather than remaining a strategic enabler [TechCrunch, August 2026] [Techticia.com, August 2026]. Investors should also watch whether Mirendil can substantiate its efficiency claims, including the reported training of Mirendil-Base by one researcher in 3.5 months with less than $10 million in compute, because those assertions are central to the company's wedge and remain only lightly corroborated in public reporting [Yahoo Finance / PR Newswire, November 2026] [pasqualepillitteri.it, October 2026].

Lightly corroborated -- This section relies on a mix of company announcement material, mainstream tech reporting, and public founder profiles; core funding facts are corroborated, while product-performance claims remain only partially verified.

Taxonomy Snapshot

Axis	Value
Stage	Series A
Business Model	SaaS
Industry / Vertical	Deeptech
Technology Type	AI / Machine Learning
Geography	North America
Growth Profile	Venture Scale
Founding Team	Co-Founders (4)
Funding	$100M+ total disclosed, approximately $1.2B [WebProNews, November 2026] [Yahoo Finance / PR Newswire, November 2026]

How the Company Got Here

Overview

PUBLIC Mirendil appears to have formed quickly around a concentrated founder bench rather than a long public buildout. The company is based in San Francisco and describes itself as a frontier lab building systems for AI research and development, a framing that is consistent with its website and company profiles rather than just financing coverage [Mirendil] [Dealroom.co]. Public materials identify Mirendil as founded in 2025, with Behnam Neyshabur as co-founder and CEO alongside co-founders Harsh Mehta, Shayan Salehian, and Tara Rezaei [Mirendil] [LinkedIn].

The public record on legal structure is thin in the supplied sources, and no state filing is cited here. What is visible is enough to place the company in San Francisco and to anchor the founding team to prior research roles at large AI labs, although those biographical details sit outside the narrow source set used for this section and are therefore left to later sections [Mirendil] [Dealroom.co].

Chronology

PUBLIC The first clear milestone in the source set is the company's 2025 founding, followed by a rapid move into large-scale fundraising in 2026. Dealroom and the company website support the existence and location of the business, while November 2026 coverage and company materials show Mirendil presenting itself as an AI R&D platform company rather than a single-model vendor [Dealroom.co] [Mirendil].

By June 2026, Mirendil had a reported $200 million seed round led by Andreessen Horowitz and Kleiner Perkins, according to later public coverage summarizing the financing history [WebProNews, November 2026]. On November 12, 2026, the company announced a $1 billion Series A led by Kleiner Perkins and Clearlake Capital, positioning that raise as capital to broaden access to frontier AI R&D [Yahoo Finance / PR Newswire, November 2026]. The chronology matters because it suggests Mirendil reached unusual funding scale within roughly a year of founding, but the underlying public evidence remains strongest for the Series A announcement and weaker for the earlier round details [Yahoo Finance / PR Newswire, November 2026] [WebProNews, November 2026].

Lightly corroborated -- Confirmed in part by the company website and Dealroom, with financing chronology supplemented by later public coverage rather than primary filings.

Product and Technology

MIXED

Mirendil is presenting itself less as a model vendor and more as an AI research automation layer. Public reporting and the company’s November 2026 funding announcement describe Lor as a private beta agent platform that can coordinate "hundreds of agents" to write code, run experiments, and provision compute over multi-day or multi-week workflows [Yahoo Finance / PR Newswire, November 2026] [WebProNews, November 2026]. That framing matters because it places the company’s product surface in workflow orchestration and research execution, not only in model performance.

The rest of the public product story is narrower, but still notable. The same announcement says Mirendil-Base is a coding model trained from scratch, and says one researcher trained it in 3.5 months using less than $10 million in compute, a claim that appears to come from the company and should be treated accordingly [Yahoo Finance / PR Newswire, November 2026]. Separately, secondary web sources describe Fern as a post-trained model with a context window of up to 10 million tokens, but those claims do not appear independently verified in the captured mainstream coverage, so they are better read as product positioning than as established technical capability [Yahoo Finance / PR Newswire, November 2026] [explainx.ai Blog].

The infrastructure picture is only partly visible in public. TechCrunch reported in August 2026 that Mirendil signed a Google Cloud deal worth more than $100 million, and separate coverage tied that arrangement to access to TPUs, Nvidia GPUs, and managed training clusters, which fits the company’s stated focus on compute-intensive AI R&D workloads [TechCrunch, August 2026] [Techticia.com, August 2026]. Beyond that, the stack should be described cautiously: Mirendil’s own web presence and social postings indicate Lor is free in private beta and available across macOS, Linux, and web, while roles in finance, company building, and founding GTM suggest the company is now adding operating and commercialization layers around a still-early technical product (inferred from job postings) [Mirendil] [Perplexity Sonar Pro Brief, November 2026].

Lightly corroborated -- Core product descriptions are corroborated by company announcement coverage and WebProNews, but several technical claims remain company-led or second-hand, with only partial independent verification.

Where the Demand Sits

Market context

PUBLIC The market matters now because Mirendil is positioned at the intersection of two spending streams that are both expanding quickly, frontier-model development infrastructure and software that automates applied AI work, but the public record on the company is still much stronger on financing than on end-market sizing [Yahoo Finance / PR Newswire, November 2026] [TechCrunch, August 2026]. In practical terms, Mirendil is not selling consumer AI access; public descriptions frame Lor as a system for writing code, running experiments, and provisioning compute for extended research cycles, which places it closer to AI R&D tooling than to horizontal chat applications [Yahoo Finance / PR Newswire, November 2026] [WebProNews, November 2026].

Because no named third-party TAM study for Mirendil’s exact category appears in the supplied research, the cleanest public proxy is an analogous market map rather than a direct size claim. The nearest adjacent markets in the source set are cloud compute for AI training and enterprise software for model development workflows. TechCrunch’s August 2026 reporting on Mirendil’s Google Cloud agreement suggests infrastructure demand is already material at the company level, with a deal described as worth upward of $100 million and tied to access to TPUs, Nvidia GPUs, and managed training clusters [TechCrunch, August 2026] [Techticia.com, August 2026]. That does not establish market size by itself, but it does indicate that the cost center Mirendil is trying to compress, model experimentation and training operations, is large enough to support nine-figure infrastructure commitments for ambitious labs [TechCrunch, August 2026].

Demand drivers in the public evidence are fairly consistent. First, model-development work is becoming more iterative and more compute-intensive, which increases the value of systems that can parallelize experimentation and manage long-running tasks [Yahoo Finance / PR Newswire, November 2026] [WebProNews, November 2026]. Second, the company’s own product framing, that one researcher can direct many agents over days or weeks, speaks to a labor-productivity pitch aimed at scarce AI research talent rather than broad workflow automation [Yahoo Finance / PR Newswire, November 2026]. Third, the founder backgrounds cited across TechCrunch, Business Insider, and founder profiles suggest the initial wedge is credible only because the product is being built by people with direct exposure to frontier-model pretraining, reasoning, and internal research tooling at Anthropic, Google DeepMind, xAI, and OpenAI-related environments [TechCrunch, August 2026] [Business Insider, February 2026] [Behnam Neyshabur CV] [tarark.com].

Adjacent and substitute markets are easier to identify than the core category, which is still loosely defined in public coverage. The closest substitutes are likely internal research platforms built inside frontier labs, cloud-native MLOps stacks, code-generation tools, and model providers that are extending into agent orchestration and experimentation workflows [SiliconANGLE, November 2026] [Yahoo Finance / PR Newswire, November 2026]. That matters because Mirendil appears to be selling an integrated layer, specialized models, agents, evaluation, and compute coordination, rather than a standalone foundation model or a generic developer copilot [Yahoo Finance / PR Newswire, November 2026]. The strategic question is whether buyers treat that integration as a new software budget line or absorb it into existing cloud and model spend.

Regulatory and macro forces cut both ways. On the supportive side, heavy investor appetite for AI infrastructure and agentic software is evident from Mirendil’s financing pace, from a reported $200 million seed in June 2026 to a $1 billion Series A announced on November 12, 2026 [WebProNews, November 2026] [Yahoo Finance / PR Newswire, November 2026]. On the constraining side, any company training models on public code and web data may face a moving legal and policy environment around data provenance, licensing, model accountability, and concentration of compute supply, particularly when access to advanced chips and cloud clusters becomes a gating factor [There's An AI For That, Retrieved 2026] [TechCrunch, August 2026]. Those are category-level pressures, not Mirendil-specific findings, but they are directly relevant to any business premised on accelerating AI R&D.

Market lens	Public evidence	What it suggests
Core category	Lor is described as a private-beta agent platform for code, experiments, and compute orchestration [Yahoo Finance / PR Newswire, November 2026]	Mirendil sits in AI R&D automation, a category still forming in public-market taxonomy
Infrastructure adjacency	Mirendil reportedly signed a Google Cloud deal worth upward of $100 million [TechCrunch, August 2026]	The underlying compute layer is already a large enterprise spend pool
Buyer proxy	Public descriptions point to AI researchers and model-development teams rather than general knowledge workers [Yahoo Finance / PR Newswire, November 2026] [WebProNews, November 2026]	Near-term demand is likely concentrated in advanced technical organizations
Substitute set	Internal lab tooling, MLOps platforms, coding agents, and cloud AI stacks are plausible alternatives [SiliconANGLE, November 2026] [Yahoo Finance / PR Newswire, November 2026]	Budget capture may depend on replacing fragmented workflows, not creating net-new demand

The table shows why market sizing remains more qualitative than quantitative in the public record. The strongest evidence is not a Gartner-style TAM number, but the fact that sophisticated buyers already spend heavily on the compute and workflow layers Mirendil is trying to compress into one system [TechCrunch, August 2026] [Yahoo Finance / PR Newswire, November 2026].

Lightly corroborated -- This section relies on named public reporting from TechCrunch, Yahoo Finance / PR Newswire, WebProNews, SiliconANGLE, and founder profiles, but it does not include a directly cited third-party TAM study for Mirendil’s exact category.

Competitive Landscape

MIXED Mirendil is positioning itself less as another model vendor and more as an automation layer for AI research itself, which places it in a different lane from broad foundation model labs even when it shares many of the same technical ingredients [Yahoo Finance / PR Newswire, November 2026] [TechCrunch, August 2026].

The public record supports a three-part map. First are the frontier model incumbents, most visibly Google DeepMind and Anthropic, which matter here because Mirendil’s founders and product framing are tied to workflows those labs already run at scale [Business Insider, February 2026] [Fortune, June 2026] [Behnam Neyshabur CV, Retrieved 2026]. Second are agent-platform challengers, where Mirendil is trying to define a category around long-running, multi-agent research execution rather than chat or code assistance, according to its Series A announcement and product descriptions [Yahoo Finance / PR Newswire, November 2026] [WebProNews, November 2026]. Third are adjacent substitutes, chiefly cloud and infrastructure providers such as Google Cloud, which can bundle training infrastructure, orchestration, and model access even if they are not direct product matches today [TechCrunch, August 2026] [Techticia.com, August 2026].

That positioning gives Mirendil one real edge today: unusually strong technical talent density for a company founded in 2025, with co-founders publicly linked to Anthropic, Google DeepMind, xAI, and OpenAI-adjacent work [LinkedIn, Retrieved 2026] [Business Insider, February 2026] [NBC News, Retrieved 2026] [tarark.com, Retrieved 2026]. The second edge is access to capital and compute, evidenced by a disclosed $200 million seed round in June 2026, a $1 billion Series A announced on November 12, 2026, and a reported Google Cloud agreement worth more than $100 million [WebProNews, November 2026] [Yahoo Finance / PR Newswire, November 2026] [TechCrunch, August 2026]. Those are meaningful advantages, but they are perishable. Talent can be copied through hiring, and compute access is valuable only if it produces a workflow customers cannot assemble from existing model APIs, internal tools, and cloud credits.

The company is more exposed on distribution than on research pedigree. No named customers, production deployments, or revenue figures are established in the available public sources, while Lor is still described as being in private beta [Yahoo Finance / PR Newswire, November 2026] [WebProNews, November 2026]. That leaves an opening for Google Cloud in particular: if enterprises prefer to buy AI R&D tooling from the same vendor providing TPUs, GPUs, and managed training clusters, the cloud platform can own the channel even without matching Mirendil feature for feature [TechCrunch, August 2026] [Techticia.com, August 2026]. Anthropic and Google DeepMind are different threats. Their advantage is not channel ownership alone, but the ability to fold internal research automation back into frontier models before an independent platform establishes a standard.

Over the next 18 months, the most plausible competitive outcome is bifurcation rather than winner-take-all. Mirendil can emerge as the winner if the market rewards an independent research operating system that helps small teams perform frontier-style experimentation without building a lab’s internal stack from scratch, which is the company’s core claim around Lor and Mirendil-Base [Yahoo Finance / PR Newswire, November 2026] [WebProNews, November 2026]. The more likely loser under the opposite condition is the standalone agent-platform category itself: if Google Cloud packages orchestration, model access, and compute procurement tightly enough, much of the buyer demand could collapse into infrastructure purchasing, leaving independent platforms with weaker pricing power and narrower room to differentiate [TechCrunch, August 2026] [Techticia.com, August 2026].

Opportunity

PUBLIC The prize here is unusually large: if Mirendil can make AI research itself more software-driven, it is aiming not at a single model launch but at the operating system for how frontier model teams build, test, and improve systems [Yahoo Finance / PR Newswire, November 2026] [TechCrunch, August 2026].

The headline opportunity is straightforward, even if the execution path is not. Public reporting describes Mirendil as building Lor, a private-beta platform that coordinates "hundreds of agents" to write code, run experiments, and provision compute over extended periods, while the company also claims internal model-development results such as training Mirendil-Base from scratch in 3.5 months with less than $10 million in compute [Yahoo Finance / PR Newswire, November 2026] [WebProNews, November 2026]. If even part of that claim set holds up in production, Mirendil is not selling another model endpoint so much as a labor-multiplication layer for AI R&D teams, which is a much larger and stickier budget line than standalone inference software [Yahoo Finance / PR Newswire, November 2026] [SiliconANGLE, November 2026]. The reason that outcome looks reachable, rather than purely aspirational, is that investors have already financed the company at uncommon scale, and Google Cloud reportedly committed to a $100 million-plus infrastructure deal that gives Mirendil access to TPUs, Nvidia GPUs, and managed training clusters, suggesting counterparties view the workload as material rather than theoretical [TechCrunch, August 2026] [Techticia.com, August 2026] [Yahoo Finance / PR Newswire, November 2026].

There are a few plausible paths to very large scale, and they do not all depend on Mirendil becoming the winning foundation-model company. One path is to become the default orchestration layer for internal AI research teams, where the buyer cares less about a single benchmark and more about compressing experimentation cycles across pretraining, evals, debugging, and infrastructure management [SiliconANGLE, November 2026] [WebProNews, November 2026]. A second path is to pair the orchestration layer with proprietary in-house models such as Mirendil-Base and Fern, turning the product into a more integrated research stack for customers that want long-context technical work and autonomous execution in one environment [Yahoo Finance / PR Newswire, November 2026] [explainx.ai Blog]. A third path is narrower but still consequential: Mirendil could become a premium platform used by a relatively small number of frontier labs, cloud partners, and model-native enterprises, where contract values are high because the product sits directly on top of compute budgets and researcher headcount [TechCrunch, August 2026] [Yahoo Finance / PR Newswire, November 2026].

Scenario	What happens	Catalyst	Why it's plausible
Research OS for frontier labs	Lor becomes a standard control plane for AI model-development workflows inside top labs and model-native enterprises	A visible production launch beyond private beta, especially if tied to a major cloud or infrastructure partner [Yahoo Finance / PR Newswire, November 2026] [TechCrunch, August 2026]	Public reporting already describes Lor as coordinating hundreds of agents over days or weeks, and Mirendil has secured unusually large funding and compute access for a company at this stage [WebProNews, November 2026] [Techticia.com, August 2026]
Full-stack AI R&D platform	Mirendil combines orchestration, coding models, long-context research models, and managed compute into one premium environment	Customer demand for fewer vendors across experimentation, coding, and training operations [Yahoo Finance / PR Newswire, November 2026]	The company has already introduced Lor, Mirendil-Base, and Fern in public materials, which at minimum shows an intent to own multiple layers of the workflow [Yahoo Finance / PR Newswire, November 2026] [explainx.ai Blog]
High-ACV infrastructure partner	Mirendil wins a concentrated set of large contracts from organizations spending aggressively on model development	The Google Cloud deal becomes a repeatable template with cloud and enterprise partners [TechCrunch, August 2026] [Techticia.com, August 2026]	A reported $100 million-plus Google Cloud agreement indicates counterparties may be willing to structure nine-figure relationships around Mirendil's compute-intensive roadmap [TechCrunch, August 2026]

The compounding logic is stronger here than in many model startups, because each product win could improve both the software layer and the economics of future development. If Lor is used to coordinate experiments, write code, manage compute, and evaluate results, then product usage can generate workflow data on which agent configurations, models, and infrastructure choices produce better research throughput, at least in principle [Yahoo Finance / PR Newswire, November 2026] [SiliconANGLE, November 2026]. That creates a plausible flywheel: better tooling attracts more advanced users, more advanced users create more demanding workflows, those workflows improve the orchestration layer, and stronger product performance justifies larger compute commitments and enterprise contracts. Early signs of that loop are still mostly company-described, but the fact that Mirendil says it uses Lor internally every day, alongside open hiring across finance, company building, and go-to-market, suggests the company is trying to turn an internal research system into a commercial platform rather than a one-off lab tool [Mirendil on X, Retrieved 2026] [Mirendil, Unknown].

The size of the win is easiest to frame by comparison to the budgets surrounding frontier AI rather than by current revenue, which is not disclosed in the sourced material. Mirendil has already announced roughly $1.2 billion in disclosed funding across seed and Series A rounds, including a $1 billion Series A on November 12, 2026, which places it among the more heavily capitalized AI infrastructure bets at an early stage [Yahoo Finance / PR Newswire, November 2026] [WebProNews, November 2026]. If the "Research OS for frontier labs" scenario plays out, and if Mirendil becomes a high-retention software and infrastructure layer attached to large recurring AI development spend, the company could plausibly support a multi-tens-of-billions valuation over time (scenario, not a forecast), because it would be selling into one of the few budgets in software still expanding at frontier-lab speed [Crunchbase News, November 2026] [SiliconANGLE, November 2026]. That upside case still needs proof on customers, product reliability, and repeatable commercialization, but the prize is clear: own the workflow where AI systems themselves get built.

Lightly corroborated -- Funding is corroborated by multiple public sources, but product capabilities and commercialization evidence remain partly company-led or based on limited independent reporting.

Sources