Measured, not claimed.
These are the models we run on hardware we control, with numbers taken from the production box on 11 November 2026. Where a number does not exist, the card says not measured rather than offering a plausible guess.
qwen3.8:27b — the on-premises workhorse
| Field | Value | How measured |
|---|---|---|
| Parameters | 27.3B, Q4_K_M GGUF | Ollama /api/ps |
| VRAM resident | 18.54 GB | size_vram |
| Throughput | 10.7 tokens/sec | 120 tokens in 11.25 s of generation |
| Cold load | 15.5 s | load_duration |
| Leaves the machine | Nothing | loopback only |
| Watts per inference | Not measured. Per-inference power telemetry is not wired yet, and we will not estimate it | |
| Benchmarks | Not run. We do not quote scores we have not produced on this box | |
Ten point seven tokens per second is fine for an email-shaped answer and too slow for the work that needs a frontier model. That gap is exactly why the hosted lane exists, and why the roadmap is capital rather than cleverness. See the Transparency Act.
ngari-ft-distilled-v5:stable — the small specialised model
| Parameters / size | 1.5B · 986 MB |
| Context | 32,768 tokens |
| Leaves the machine | Nothing |
| Throughput, watts, evals | Not measured on this box |
Hosted models — not ours
deepseek-flash, deepseek-v4-pro (DeepSeek, China) · kimi-k3.1-pro
(Moonshot, China) · qwen3.8-flash, qwen3.8-max (Alibaba, China). Retention is
provider_terms. We publish what we send, and we do not claim to measure their throughput.
Source and measurement method: the Transparency Act and the model-card document in the repository. If a number here disagrees with the box, the box is right and this page is a defect — write to support@ngari.ai.