Mirendil announced Lor, built "to do the work of an entire AI lab." Its agents "write code, run experiments, and iterate on the results for days, on any machine, hundreds at a time." Mirendil's pitch is that with agents doing the hands-on work, "one person can take on what used to need a team, and a team what used to need an institution," while people make the calls about what to try and what to trust. Mirendil says it runs on Lor every day. Lor is in private beta.
What Lor does
- Hundreds of agents in parallel: an agent can split one job across dozens of agents, and hundreds across a user's projects. The examples given are reading thousands of agent traces for failures, running hundreds of experiments and building training environments from real pull requests. Users can follow each agent's progress and stop any of them at any time.
- Their own compute: Lor comes with its own compute. Agents request nodes, set them up, run the work and release them when done, with no cloud account or setup. Mirendil says that is what lets Lor handle AI development end to end: an experiment, a training run or an eval sweep, with the agent spreading the work across nodes and handing them back when results are in.
- Days or weeks of work: Lor is built for durable, asynchronous work, and agents notice when something breaks and unblock themselves. Mirendil's examples are turning raw user feedback into shipped fixes, babysitting training runs overnight and keeping automations going for weeks.
- Any machine: agents run in the cloud and can work on Lor-managed cloud computers, a user's laptop or any other machine. Mirendil describes setting up a team of agents from a laptop, sending training runs to different GPU nodes and following it all from a phone.
The models
Lor runs frontier models alongside open-source ones and Mirendil's own. The listed lineup is 25 models:
- Mirendil: Fern, the first of its own models, which holds up to 10M tokens of context, "enough for days of an agent's work at once."
- Anthropic: Claude Opus 5.5, Claude Fable 5.1, Claude Sonnet 5.5, Claude Haiku 5.5, Opus 5 and Fable 5.
- OpenAI: GPT-6 Astra, GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Sol and GPT-6 Luna.
- Google: Gemini 4 Argon.
- xAI: Grok 4.7 and Grok 4.6.
- Meta: Muse Spark 1.3 and Muse Wonder.
- Z.ai: GLM-5.3, GLM-5.3-Flash, GLM-5.2 and GLM-4.7-Flash.
- Moonshot AI: Kimi K3.1 Pro and Kimi K3.
- DeepSeek: DeepSeek 4.1 Flash and DeepSeek 4 Pro.
Mirendil-Base
Mirendil says the approach has already been tried on itself. A single Mirendil researcher recently pretrained Mirendil-Base, "a competitive coding model," with agents handling almost all of the engineering, a project that once needed an entire team and that one person drove in three and a half months. Mirendil's stated aim is "a frontier AI lab for every scientist," so that building AI is no longer the bottleneck.
Availability
Lor is invite-only, with a waitlist, and free during the private beta. It runs on macOS, Linux and the web, with Windows, iOS and Android coming soon. The platform claims are Mirendil's own.