Catalogue

Models

The models actually served, with their parameters, quantisation, context window, memory footprint, licence and measured speed. Every number has a date and a method.

The catalogue

Served as	Artifact	Parameters	Quantisation	Context	Memory	Licence	Measured speed
`mh-sovereign-small`	Qwen3.5-4B Q4_K_M (unsloth GGUF)	4B	`Q4_K_M`	16,384 tokens	4.5 GB · 2.74 GB file	Apache-2.0	38–40 tok/s GPU · ~15 tok/s CPU 2026-10-12 — single stream, bench2
`mh-sovereign-reason` default — alias mh-sovereign	Qwen3.5-9B UD-Q4_K_XL (unsloth GGUF)	9B	`UD-Q4_K_XL`	32,768 tokens	9.0 GB · 5.97 GB file	Apache-2.0	24.6 tok/s GPU · TTFT 0.3–0.9 s · 14/14 task classes 2026-10-12 — single stream, thinking off, bench2
`mh-sovereign-lingua`	Gemma-4-12B QAT Q4_K_XL (unsloth GGUF)	12B	`QAT Q4_K_XL`	20,480 tokens	9.5 GB · 6.72 GB file	Apache-2.0	17.2 tok/s GPU · 14/14 task classes 2026-10-12 — single stream, bench2

`mh-sovereign` is the default alias and resolves to `mh-sovereign-reason`.

Cloud figures are projected until measured on cloud hardware; every number on this page was measured on one local machine with one user.

Known limits

Numbers need thinking

Measured

Tasks involving money, prices or quantities need thinking enabled on the 9B model (budget 1,024 tokens); without it, arithmetic errors are more frequent.

Estonian

Measured

No served model writes publishable Estonian. Drafts need a native editor — this site's own Estonian pages carry that notice.

Latvian

Measured

Usable only from the 9B model and with editing.

One model resident at a time

Measured

One model is loaded at a time and unloaded after 120 seconds idle. A request for another model costs a load; a dedicated instance keeps your models resident.

Status labels — MEASURED — exists and was measured on the stated date by the stated method. BUILT — exists and works, not yet measured under load. PLANNED — designed, not yet built. A line without a label is not a claim of this platform.

Method

Measured on 2026-10-12 on NVIDIA GeForce RTX 5070 12 GB · AMD Ryzen 7 7700 · 63 GB RAM · Windows 11 Pro, runtime llama.cpp b11559, CUDA 13.4 (Windows binaries); one user, GPU shared with media production under a governor (GPU norm 34 %), free heavy.lock windows. Evidence folders: media-holding-sovereign-ai/evidence/bench2, media-holding-sovereign-ai/evidence/soak.

Cloud numbers are projected until a cloud node is measured under the same method. A dedicated instance on a 24 GB GPU is expected to run the 9B and the 12B model resident at the same time; that expectation is a projection, not a measurement.

Measured through the platform

Measured

On 2026-10-13 one real single-user request went through the Sovereign AI Cloud control plane to Sovereign AI (mh-sovereign-reason (Qwen3.5-9B UD-Q4_K_XL), RTX 5070, GPU norm 34 %): 37 tokens generated at 25.8 tokens per second; prompt of 21 tokens processed in 297 ms; model load 2.1 s. Evidence: sovereign-ai-cloud/evidence/real_inference_single_user_20261013.json.

Who developed the models

Model	Developer	Role
Qwen3.5 (4B, 9B)	Alibaba — Qwen team	Model developer
Gemma 4 (12B)	Google DeepMind	Model developer
GGUF quantisations	Unsloth (GGUF quantisations)	Quantised artifacts
llama.cpp	ggml-org (llama.cpp)	Inference runtime

Under the EU AI Act the obligations of general-purpose AI model providers lie with the model developers; Media Holding does not modify the weights of the models it serves. The AI Transparency Notice states what this means for you as a deployer.