This dossier draws on 44 selected evidence rows across 39 independent source domains, captured through 2026-10-30: press coverage, company posts, investor databases, job listings and interviews. Mirendil is a San Francisco frontier-AI lab pursuing "self-accelerating" — recursively self-improving — AI. The timeline below separates verified events, company claims, and reported-but-unconfirmed items.
Privately held AI research lab at 275 Battery Street, San Francisco. Founding is dated variously to December 2025 (founders' Anthropic departure) and early/2026; the public launch came June 2026 — the discrepancy is unresolved in the sources. linkedin.com tmtpost.com
Build models and tools that conduct AI R&D themselves — and over time help build better successor models — to accelerate scientific discovery in fields like biology, materials science and drug discovery, and to democratize frontier AI R&D. index.fame.so haystackapp.io
Behnam Neyshabur (CEO; ex-Anthropic, Senior Staff Research Scientist at Google, co-led Gemini, created Minerva), Harsh Mehta (CTO; initiated Anthropic's automated pretraining R&D project), Shayan Salehian (early ML engineer at xAI), Tara Rezaei (MIT graduate, early student researcher at OpenAI). Backgrounds span Anthropic, Google DeepMind, xAI and OpenAI. alphaxiv.org investing.plus
$200M seed completed late June 2026 at a reported $1B valuation, co-led by Andreessen Horowitz and Kleiner Perkins with Nvidia participating. Later reports describe talks to raise up to $1B at a $5B valuation — negotiations only, not completed financing. citations.press cbinsights.com
Company claim: Mirendil-Base, a 100B–500B-parameter coding base model, pretrained by one researcher directing agents in 3.5 months for under $10M all-in, built on libraries such as PyTorch/JAX and public data, agents writing over 99% of the code. Larger Base variants (up to a planned 2T–10T "XL") are in training per technical coverage. Baseten, its inference partner, reported the lowest bits-per-byte of any model it tested — a partner result, not an independent eval. mindpattern.ai twstalker.com
Fern 0.1 (2026-10-26): post-trained Base model claiming a 10M-token window, for auditing training data, debugging failed runs from logs and reading long agent traces; early-access request only, no pricing. Lor (2026-10-28): an agent platform meant to do the work of an AI lab, private beta by invite, free during beta. No public API or generally available product. wearethefrontier.org mirendil.com
Signed multi-year Google Cloud agreement worth more than $100M (reported August 2026): TPUs, Nvidia GPUs and managed training clusters — committed before any public product, benchmark or revenue. bitroot.org rwg3ewym.mule.page
Open roles span pretraining, post-training/RL, evaluations, kernels, distributed infrastructure, data and a first design hire. Some technical roles advertise $300,000–$400,000 base plus equity; at least some are five days a week in San Francisco with long, high-intensity in-person days. jobs.nonlinearproject.com linkedin.com
No confirmed customer list, meaningful revenue disclosure, durable business model, public API, generally available product or pricing appears anywhere in the selected evidence. An API-readiness index notes no developer portal, SDK or documentation surface as of mid-2026. providers.apievangelist.com
Covered as part of the "neo-labs" / RSI startup wave alongside peers; critics note no clear differentiation yet from general coding agents, and that $200M shows it can pay experiment bills, not that its models research better — "capital ahead of evidence". huxiu.com app.dealroom.co
| Published | Topic | Finding | Source |
|---|
Method: 44 selected evidence rows on Mirendil from 39 independent source domains, captured as of 2026-10-30, covering company profile, people, technology, funding, infrastructure, hiring and risk commentary. Dollar figures are as reported; valuations and the $5B round are reported figures, not audited. Benchmark numbers are company-run (20 examples per RULER result). Long findings and caveats are trimmed in the table for space; follow each source link for the full text. Some sources are aggregators; official and original reporting is preferred where rows conflict.