One Box, Four Models, Matched to the Job
Reeve uses a small set of open models, each picked for one job and measured on our own box. Ollama runs them locally. Only one model needs to be in memory at a time, and switching takes about one to four seconds.
Last updated| Job | What It Does for You | Model | Licence | Memory | Speed |
|---|---|---|---|---|---|
| Router | Sorts and classifies questions, alerts and tickets | Qwen3 8B | Apache 2.0 | 7.5 GB | ~100 tokens/s |
| Main agent | Answers, plans and calls your tools | gpt-oss 20B | Apache 2.0 | 12.9 GB | ~139 tokens/s |
| Coder | Builds and maintains connectors to your systems | Qwen3-Coder 30B | Apache 2.0 | 20.1 GB | ~123 tokens/s |
| Vision | Reads a screen to work tools that have no API | Qwen3-VL 8B | Apache 2.0 | 7.8 GB | ~98 tokens/s |
| Search | Indexes your documents so answers can cite them | nomic-embed-text | Apache 2.0 | small | — |
Memory is measured with a 16K-token context window. Speed is the median generation rate on the reference box. Alert triage stays with a person for now: in our tests no model beat a simple baseline, so Reeve prepares the evidence and a person decides.
Every Model Earns Its Place
It Must Run Here
It must support tool calling, ship real local weights rather than a cloud-only version, and come from the official library.
A Business-Safe Licence
Apache 2.0 or MIT only, read from the model's own card. Models with custom licences are dropped.
Measured, Not Marketed
Test suites are written before any model runs, scored by code, three repeats each. Published benchmark tables don't count.
Production Changes Last
A new model replaces the current one only when it wins on our tests, and the switch ships through staged updates.
What It Takes to Run
Every Reeve system runs on AMD Radeon and Instinct GPUs with AMD’s ROCm software. Every standard model runs on a single workstation-class box, and Size S is the reference box we measured on. Bigger sizes add GPU memory for more people and agents.
| Part | Size S Reference · Measured | Size M | Size L | Size XL |
|---|---|---|---|---|
| GPU | AMD Radeon RX 7900 XTX, 24 GB | AMD Radeon PRO W7900, 48 GB, or 2× AMD Radeon AI PRO R9700, 64 GB | 2× AMD Radeon PRO W7900, 96 GB, or 4× AMD Radeon AI PRO R9700, 128 GB | Two or more Size L servers, or an AMD Instinct MI300X server (192 GB per accelerator) |
| CPU | AMD Ryzen 7 5800X3D | AMD Ryzen 9 or Threadripper, 16 cores | AMD Threadripper PRO or EPYC, 24–32 cores | AMD EPYC, per server |
| Memory | 32 GB | 64–128 GB | 256 GB | 256 GB+ per server |
| Storage | 1 TB NVMe | 2 TB NVMe | 4 TB NVMe, mirrored | Mirrored NVMe, per server |
| Software | Ubuntu Linux, ROCm, Ollama | Ubuntu Linux, ROCm, Ollama | Ubuntu Linux, ROCm, Ollama or a high-throughput engine | Ubuntu Linux, ROCm, a high-throughput engine |
Only Size S has been measured. Sizes M to XL are planned builds on the same AMD platform. How the sizes compare ↓
Sized for More People and More Agents
Every box runs the same Reeve software and the same controls. What changes as you grow is how much GPU memory it has, which sets how many models stay loaded and how many people and agents it can serve at once.
Office Box
One team getting started
- GPU memory
- 24 GB, one card (reference: AMD Radeon RX 7900 XTX)
- CPU / RAM
- 8 cores · 32 GB
- Storage
- 1 TB NVMe
- Form
- Desktop tower
ModelsOne model in memory at a time, swapped by job in about 1–4 seconds.
RecoveryEncrypted backups and tested restores.
Team Workstation
Several teams, plus IT and security work
- GPU memory
- 48–64 GB: AMD Radeon PRO W7900, or two AMD Radeon AI PRO R9700
- CPU / RAM
- 16 cores · 64–128 GB
- Storage
- 2 TB NVMe
- Form
- Workstation tower
ModelsAt 48 GB, router, main agent and vision stay loaded together (about 28 GB) with room for several conversations at once, and the coder loads on demand. At 64 GB, all four can stay loaded.
RecoveryBackups, plus an optional standby box.
Department Server
Organization-wide use at one site
- GPU memory
- 96–128 GB: two AMD Radeon PRO W7900, or four AMD Radeon AI PRO R9700
- CPU / RAM
- 24–32 cores · 256 GB
- Storage
- 4 TB NVMe, mirrored
- Form
- Rack server, redundant power
ModelsAll four standard models stay loaded (about 48 GB), with headroom for many requests at once, or a larger model for harder work.
RecoveryStandby box kept in step.
Multi-Server
Larger organizations and multiple sites
- GPU memory
- 192 GB+: two or more Size L servers, or AMD Instinct MI300X
- CPU / RAM
- Per server, as size L
- Storage
- Per server, mirrored
- Form
- Two or more rack servers
ModelsEvery model loaded on every server. Work is shared between them, and one takes over if another fails.
RecoveryAutomatic failover between servers.
People and agent counts are planning estimates, not measurements. Your box is sized during the readiness assessment. Measured figures will replace them after load testing.
People at the Same Time
Capacity depends on how many people are asking at once, not total headcount. Fifty staff rarely means fifty questions in the same second.
Models Kept Loaded
More GPU memory keeps more models ready, so there’s no wait while the box swaps from one job to another.
Long Documents
Reading a 300-page contract uses far more memory than a short question. Heavy document work needs more room.
Agents Working in the Background
Most agents sit idle until called. Busy agents, like alert triage during an incident, compete with people for the same memory.