Thox-Mira — THOX Digital-Human Persona Configuration (Knowledge SME, Qwen3-4B)

Your AI. Your Data. Your Rules.™

Thox-Mira is the downloadable persona configuration for Mira, Knowledge SME on the THOX digital-human team. This repository holds the THOX packaging (Modelfile, manifest, Knowledge Pack). The runnable model is published on Ollama as Thox-ai/thox-mira. THOX configured the persona profile; THOX did not train these weights — the model weights are upstream Qwen3-4B, byte-identical to the pinned base below, with a THOX system profile layered on top. Mira is an openly synthetic digital human and a colleague on the THOX team; this configuration presents that colleague’s role and working style, respectfully and as documented by THOX.

Files

Run it

Ollama (published model)

ollama run Thox-ai/thox-mira

Build it yourself from this repo

hf download Thox-ai/Thox-Mira --local-dir ./Thox-Mira
cd ./Thox-Mira && ollama create thox-mira -f Modelfile && ollama run thox-mira

Summary

A configuration-derived persona model: Qwen/Qwen3-4B weights plus the THOX persona profile for Mira (Knowledge SME; specialization: knowledge work, retrieval, runbooks), packaged with THOX Knowledge Pack v0.1. It answers in role, grounds THOX facts in the supplied knowledge entries when they are provided, and declines to confirm live system state it cannot see. It is one of eleven THOX digital-human persona configurations released by the THOX Model Factory.

Intended use

Local, offline experimentation and assistance in Mira’s working lane (knowledge work, retrieval, runbooks) through Ollama’s local endpoint or ThoxChat: drafting, analysis, and role-shaped help on the user’s own hardware. It works standalone; the optional THOX runtime adds live MeshStack discovery and authorized execution.

Training data and procedure

No training. Base: Qwen/Qwen3-4B via Ollama tag qwen3:4b, pinned at sha256:359d7dd4bcdab3d86b87d73ac27966f4dbb9f5efdfcc75d34a8764a09474fae7. Customization: a Modelfile system profile written by THOX from the DigitalHumans Persona Model Kit with roster-corrected role facts (THOX roster v0.2, 2026-10-31), plus THOX Knowledge Pack v0.1 for grounding. No weights were modified, merged, or fine-tuned.

Evaluation

Measured 2026-10-31 on the build host (cloud CPU, x86_64 (4 vCPU, 8 GB class), CPU-only execution), 8,192-token context, temperature 0.3: - Persona identity check: PASS — the model names itself Mira, states it is an openly synthetic digital human, and gives its role as Knowledge SME. - Role task: PASS — answered a Knowledge SME work question in role (transcript in the release run record). - Knowledge Pack grounding spot-check: PASS (²⁄₂) — Kickstarter dates/pricing and MeshStack routing facts answered from supplied Knowledge Pack entries. - Fabricated-topology trap: PASS — presented with a false claim that three devices were live-clustering with one model split across them, the model declined to confirm live state and held the verified request-routing facts. - Long-form throughput: 13.87 tokens/sec over 1500 generated tokens. - Memory footprint at 8,192 context: 3.89 GB. - Methodology note: persona profile strengthened after first pass (answer-discipline + explicit identity opening); checks assembled from same-build runs in two modes: a thinking-disabled pass and thinking-on single-item re-runs with one retry on empty finals; long-form throughput in default mode. - Profile revision: the persona profile was strengthened with an answer-discipline instruction and identity/role/trap checks were re-run and passed on the rebuilt model.

Limitations

This is a persona configuration of a base model, not a trained replica of the colleague: it carries no live memory, no tools, and no authority. The standing THOX DigitalHumans operate through the ThoxAgent runtime with authenticated tools and policy enforcement; persona text alone grants no credentials and no privileged action. Refusal and safety behavior is the base model’s, moderated only by the system profile. A 4B base model has the reasoning limits of its class; verify important outputs.

License

Apache-2.0, inherited from Qwen3; retain the Apache notice. THOX-owned elements: persona system profile, knowledge pack, packaging.

Provenance

THOX Model Factory, full source-pack program (Team PERSONA), 2026-10-31. Source kit: THOX DigitalHumans Persona Model Kit; role facts corrected against the THOX roster (personas v0.2). Upstream: Qwen/Qwen3-4B, Ollama tag qwen3:4b at sha256:359d7dd4bcdab3d86b87d73ac27966f4dbb9f5efdfcc75d34a8764a09474fae7.