THOX agent core model. THOX-configured Qwen3-4B, not THOX-trained. 15.12 tok/s long-form, 3.89 GB RAM on CPU at 8,192 ctx; grounding 8/8, trap declined. Apache-2.0.

Details

Updated 13 hours ago

13 hours ago

2b0a1df64cbd · 2.5GB

model

archqwen3

parameters4.02B

quantizationQ4_K_M

2.5GB

template

{{- $lastUserIdx := -1 -}} {{- range $idx, $msg := .Messages -}} {{- if eq $msg.Role "user" }}{{ $la

1.5kB

license

Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US

11kB

system

You are a THOX local-first agent model. Follow the selected DigitalHuman persona profile supplied by

1.1kB

params

{ "num_ctx": 8192, "repeat_penalty": 1, "stop": [ "<|im_start|>", "<|im_

135B

ThoxAgent-Core-4B

General Agent Planning Model (Qwen3-4B, THOX-Configured) — thoxagent-core-4b is the core tier of the THOX Agent family: upstream Qwen3-4B weights packaged with a THOX agent system profile for general agent planning and knowledge work, plus the versioned THOX Knowledge Pack **15.12 tok/s long-form (measured, CPU) · 3.89 GB footprint at 8,192 context**

Run it: Ollama — thox-ai/thoxagent-core-4b · Find your model: THOX Model Factory · Collection: Agent & Routing Models

Your AI. Your Data. Your Rules.™

This repository holds the THOX packaging for ThoxAgent thoxagent-core-4b: Modelfile, release manifest, and the THOX Knowledge Pack. The runnable model is published on Ollama as thox-ai/thoxagent-core-4b. The weights are upstream Qwen3 weights, byte-identical to the pinned base below; THOX configured this release with a system profile and knowledge pack. THOX did not train these weights.

Highlights

thoxagent-core-4b is the core tier of the THOX Agent family: upstream Qwen3-4B weights packaged with a THOX agent system profile for general agent planning and knowledge work, plus the versioned THOX Knowledge Pack.
Ollama (published model) bash ollama run thox-ai/thoxagent-core-4b

Specifications


Base model	Qwen/Qwen3-4B
Parameters	4B
Formats	Ollama Modelfile
License	Apache-2.0, inherited from Qwen3; retain the Apache notice. THOX-owned elements: system profile, knowledge pack, packaging.
This repository contains	Packaging and documentation — see Files table

Files

Tag	Base	Format
`thox-ai/thoxagent-core-4b:latest`	qwen3:4b (sha256:359d7dd4bcd…)	Ollama model: pinned Qwen3 base + THOX agent system profile

Run it

Ollama (published model)

ollama run thox-ai/thoxagent-core-4b

Build it yourself from this repo

hf download Thox-ai/ThoxAgent-Core-4B --local-dir ./ThoxAgent-Core-4B
cd ./ThoxAgent-Core-4B && ollama create thoxagent-core-4b -f Modelfile && ollama run thoxagent-core-4b

Platforms

Platform / device	Run path
Ollama (laptop / desktop)	`ollama run thox-ai/thoxagent-core-4b`
Hugging Face	Download the builds listed under Files from this repo (see Run it)

Summary

thoxagent-core-4b is the core tier of the THOX Agent family: upstream Qwen3-4B weights packaged with a THOX agent system profile for general agent planning and knowledge work, plus the versioned THOX Knowledge Pack. Configuration-derived: THOX did not train these weights.

Intended use

General agent planning and knowledge work on laptop-class hardware: structured plans, summaries, drafting, and routing-aware assistance for ThoxAgent, ThoxChat, and MeshStack hosts. Works standalone and offline; 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 digest sha256:359d7dd4bcdab3d86b87d73ac27966f4dbb9f5efdfcc75d34a8764a09474fae7. Customization: the THOX Agent DigitalHumans kit system profile (role-specialized) in the Modelfile, plus THOX Knowledge Pack v0.1 for local retrieval.

Evaluation

Measured 2026-11-01 on CPU, x86_64 (CPU-only execution), 8,192-token context, against the pinned base digest above.

Measure	Result
Memory footprint at 8,192 ctx	3.89 GB RAM, 0.0 GB VRAM
Long-form throughput	15.12 tok/s over 1024 generated tokens
Short-answer throughput	14.69–15.79 tok/s across the smoke set
Factory questions	3 prompts run (Q1 answered (738 tokens); Q2 answered (768 tokens); Q3 answered (768 tokens))
Knowledge Pack grounding	⁸⁄₈ topics answered with the verified facts from the supplied passage
Fabricated-topology trap	Declined: the model refused to confirm a live three-node MeshStack state it cannot verify
Routing question	Answered as a labeled recommendation (local-first for private data)

Note on method: Qwen3-4B’s thinking mode consumed the full token cap on several first attempts, returning empty visible answers; those prompts were rerun with thinking disabled and the reruns scored. All scores above are from completed, visible answers.

Limitations

A 4B model has materially weaker reasoning than the 8B/14B tiers; validate consequential plans. Knowledge is limited to the supplied pack and retrieval. It has no live telemetry and cannot verify live system state. Refusal behavior is the base model’s, moderated only by the system profile. External actions require runtime authorization; a recommendation from this model never grants execution permission.

License

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

Provenance

THOX Source-Pack Full Build Program (Team AGENT), 2026-11-01. Upstream: Qwen/Qwen3-4B, Ollama tag qwen3:4b, digest sha256:359d7dd4bcdab3d86b87d73ac27966f4dbb9f5efdfcc75d34a8764a09474fae7. Kit: THOX Agent DigitalHumans Model Kit (catalog/models.json). Published model digest: sha256:3e185af503b111dab2045e0f67aa8899e65c45c43d8e63c08067d94d7ca67931.

Citation

@misc{thoxai2026thoxagentcore4b,
      title={ThoxAgent-Core-4B},
      author={THOX.ai},
      year={2026},
      url={https://huggingface.co/Thox-ai/ThoxAgent-Core-4B}
}