# MIRASTACK LABS > MIRASTACK LABS runs open models on its own GPUs and sells capacity by time, not by the token: for AI factories, dedicated endpoints, teams and developers. MIRASTACK LABS PRIVATE LIMITED (Hyderabad, India) hosts and runs open-weight models on its own GPUs and sells that capacity by time, not by the token. Four offers: - AI Factory (IaaS): an AI factory in your own data centre, designed, brought live and run by us; air-gapped first. https://www.mirastacklabs.ai/ai-factory/ - infinitetokens (dedicated, PaaS): training clusters and inference reserved on your own endpoint for a scoped project and its term. https://www.mirastacklabs.ai/dedicated/ - unlimitedtokens (teams, SaaS): per-seat access to open models for a whole organisation, billed by time. https://www.mirastacklabs.ai/teams/ - endlesstokens (individuals): a self-serve time pass for developers, with unlimited tokens on open models, for any OpenAI- or Anthropic-compatible client. https://www.mirastacklabs.ai/individuals/ No prices are published; every agreement is shaped to the workload. ## miracache miracache is open-source (Apache-2.0) shared KV-cache memory for LLM inference, made of servers you already own and reached over RDMA. The Markdown files below are plain-text versions of the page and its docs. - [miracache (Markdown)](https://www.mirastacklabs.ai/miracache.md): the product page as Markdown. - [Quick start (Markdown)](https://www.mirastacklabs.ai/miracache/quickstart.md) - [FAQ (Markdown)](https://www.mirastacklabs.ai/miracache/faq.md) - [miracache (HTML page)](https://www.mirastacklabs.ai/miracache/): miracache turns the DRAM of older servers into a shared memory tier for LLM inference over RDMA. Bring your DDR4 hardware and run the first tests with us. - [GitHub repository](https://github.com/mirastacklabs-ai/miracache): source, README and releases. - [All miracache text in one file](https://www.mirastacklabs.ai/llms-full.txt) ## Company - [Company | MIRASTACK LABS](https://www.mirastacklabs.ai/company/): MIRASTACK LABS Private Limited is based in Hyderabad, India. We run AI capacity on open models and build open tooling for sovereign infrastructure. - [Contact | MIRASTACK LABS](https://www.mirastacklabs.ai/contact/): Talk to MIRASTACK LABS. Tell us the workload, the number of people or systems, and where it must run, and we will shape an agreement around it. - [Privacy notice | MIRASTACK LABS](https://www.mirastacklabs.ai/privacy/): What personal data the MIRASTACK LABS website processes, why, who processes it for us, how long we keep it, and the rights you have over it. - [Build an AI Factory in Your Data Centre | MIRASTACK LABS](https://www.mirastacklabs.ai/ai-factory/): Build an AI factory in your own data centre, on your hardware or ours. We design it (Day 0), bring it live (Day 1) and run it (Day 2). Air-gapped first. - [On-Prem Sovereign AI, India and Global | MIRASTACK LABS](https://www.mirastacklabs.ai/sovereign-ai/): Run sovereign AI on premises, air-gapped, on your hardware or ours. Covers India: the DPDP Act 2023 and in-country deployment. Built in Hyderabad. - [infinitetokens: dedicated LLM endpoint | MIRASTACK LABS](https://www.mirastacklabs.ai/dedicated/): infinitetokens: infinite tokens for a finite time. Training clusters and inference reserved on your own endpoint for 3, 6, 9 or 12 months. Never per token. - [unlimitedtokens: flat-rate LLM API | MIRASTACK LABS](https://www.mirastacklabs.ai/teams/): unlimitedtokens is a flat-rate LLM API for organisations: per-seat access to open models, billed by time and never per token. Agreements from 60 seats. - [endlesstokens for individuals | MIRASTACK LABS](https://www.mirastacklabs.ai/individuals/): endlesstokens is a self-serve time pass for developers: unlimited tokens on open models, for any OpenAI- or Anthropic-compatible client and coding agent. - [Technology | MIRASTACK LABS](https://www.mirastacklabs.ai/technology/): How MIRASTACK LABS runs time-based AI: our own GPUs, one gateway, open-source serving, and capacity measured with real agent traffic before it is sold. - [Open source | MIRASTACK LABS](https://www.mirastacklabs.ai/open-source/): Open-source work by MIRASTACK LABS: the Isolation Forest processor for OpenTelemetry, Cosmolet, the Telegen eBPF agent and miracache for LLM inference. - Email: hello@mirastacklabs.ai - LinkedIn: https://linkedin.com/company/mirastack-labs - GitHub: https://github.com/mirastacklabs-ai ## Optional - [Blog | MIRASTACK LABS](https://www.mirastacklabs.ai/blog/): Guides and engineering notes from MIRASTACK LABS: AI capacity, how LLM inference is served, model design, and the open-source tools we build. - [Case studies | MIRASTACK LABS](https://www.mirastacklabs.ai/case-studies/): Case studies from MIRASTACK LABS: real workloads on our own GPUs, measured and explained, grouped by who they are for. Individuals, teams and more. - [A Third of a Billion Tokens, One Pass | MIRASTACK LABS](https://www.mirastacklabs.ai/case-studies/individuals/a-third-of-a-billion-tokens-one-daily-pass): One ₹999 Daily pass moved a third of a billion tokens, 91% from cache, on one GPU that used at most 3 of its 16 sessions. How we measured it. - [mirafleet: AI Agents Without a Bastion | MIRASTACK LABS](https://www.mirastacklabs.ai/blog/ai-coding-agents-without-a-bastion): Keep code and credentials off laptops without a jump host: the model thinks on the laptop, keys stay on central, and every tool call runs on runners. - [Unlimited Tokens vs Per-Token Pricing | MIRASTACK LABS](https://www.mirastacklabs.ai/blog/unlimited-tokens-vs-per-token-pricing): Per-token billing turns every agent loop into a cost decision. See how flat, time-based LLM capacity works for people and teams, and when metered wins. - [Infinite Tokens for a Finite Time | MIRASTACK LABS](https://www.mirastacklabs.ai/blog/infinite-tokens-for-a-finite-time): Reserve training clusters and inference on your own endpoint for 3, 6, 9 or 12 months. How infinitetokens fits bounded model work and when it does not. - [What Is an AI Factory? Day 0, 1 and 2 | MIRASTACK LABS](https://www.mirastacklabs.ai/blog/what-is-an-ai-factory): What an AI factory is, layer by layer from GPU to site, and how to design (Day 0), build (Day 1) and operate (Day 2) one on your hardware or ours. - [On-Prem Sovereign AI, Globally and in India | MIRASTACK LABS](https://www.mirastacklabs.ai/blog/on-prem-sovereign-ai-india): What sovereign AI means, how on-premises and air-gapped compare, and what Indian organisations should know about the DPDP Act 2023 and in-country AI. - [LLM Inference Parallelism, Explained | MIRASTACK LABS](https://www.mirastacklabs.ai/blog/llm-inference-parallelism): An animated guide to LLM inference parallelism: data, tensor, pipeline and expert parallelism, and the attention designs (MHA, GQA, MQA, MLA) behind them. - [Next-Concept Modeling (NCM), Explained | MIRASTACK LABS](https://www.mirastacklabs.ai/blog/next-concept-modeling): How a language model can write a paragraph in ~12 parallel steps instead of hundreds: a coarse-to-fine hierarchy of discrete codes, one level at a time.