Stoa Docs

Appearance

Hardware sizing ​

Stoa runs on your hardware, in your data center. Picking the right machine matters: the model you can run, the number of users you can support, and how snappy chats feel are all decided by what you put underneath. This page summarizes the three sizing tiers we recommend, plus a few notes on multi-GPU setups and the current GPU market.

The three tiers ​

Most deployments fall into one of three buckets. Pick the one that matches the user count and document volume you expect in the first 12 months.

Entry tier: $2,500 to $6,000 ​

A single mid-range GPU (think a current-gen consumer card with 16 to 24GB of VRAM), 64 to 128GB of system RAM, and a modern multi-core CPU. This tier comfortably runs an 8B or 9B parameter model from the Llama, Qwen, or Mistral families and supports 5 to 10 concurrent users in a small team or pilot.

Best for: pilots, small teams, single-department deployments, proofs of concept that need to run on real data.

Growth tier: $15,000 to $22,000 ​

An NVIDIA A6000 or two consumer GPUs pooled into one host, 256GB of RAM, and enterprise-grade NVMe storage. This tier runs larger 30B to 70B parameter models, handles bigger knowledge bases, and supports 25 to 35 concurrent users.

Best for: a department or business unit, KB collections in the tens of thousands of documents, mixed chat plus deep-research workloads.

Scale tier: $55,000 to $90,000+ ​

An NVIDIA A100 or H100 (or multiple), 512GB+ of RAM, and high-throughput storage. This tier runs the largest open-source models, handles long-context research jobs in parallel, and supports 50 to 100+ concurrent users.

Best for: organization-wide rollouts, regulated environments with heavy audit and research workloads, deployments that need to run state-of-the-art open-source models locally.

Multi-GPU and GPU pooling ​

Stoa runs on multi-GPU systems, and it can pool GPUs from existing machines into a single host. If you already have a workstation with two consumer cards, or a server with several mid-range GPUs sitting idle, your delivery engineer can configure Stoa to use them as one inference pool. This is the cheapest path from "entry tier" to "growth tier" for organizations that already own GPU hardware.

TIP

Mixing GPU generations works, but the slowest card sets the pace for any model spread across them. If you mix cards, group similar ones together.

RAM and GPU pricing reality ​

Both system RAM and GPU prices have climbed sharply in the last 18 months and are still moving. The dollar figures above are good planning estimates, not quotes. We refresh the public sizing guide regularly so it tracks current market rates. Get a fresh quote before committing to a build.

Open models and commercial APIs ​

Stoa supports both:

  • Open-source models running locally on your hardware: Llama, Qwen, Mistral, and DeepSeek family models in a range of sizes. This is the default and is what almost all customers use, because it keeps every prompt and document inside your network.
  • Commercial APIs (OpenAI, Anthropic, others) when your policy allows it. Some teams want a local model for sensitive data and a commercial API for low-sensitivity tasks. Stoa lets you wire both up and route per-agent.

For air-gapped or otherwise locked-down sites, only the local-models path is in scope. See Updates & on-site delivery for how those deployments are kept current.

Storage planning ​

Storage is independent of GPU sizing. Plan for at least 100GB to start, more if you intend to ingest large document libraries. KB indexes are roughly 1.5 to 2x the size of the source documents, plus headroom for the conversation history and audit log. See Backups for what needs to be preserved long-term.

Get the full sizing guide ​

The full sizing guide, including sample bills of materials, recommended GPUs by current price, and concrete reference builds, lives on our marketing site:

See the full hardware sizing guide on stoa.microapps.io

Next ​

Stoa, a MicroApps product.

Copyright © Stoa & MicroApps