Thox-ai/ thox-ios-open-4b:latest

8 yesterday

iOS open chat configuration (profile-only, not a fine-tune). THOX-configured Qwen3-4B, not THOX-trained. 15.34 tok/s long-form, 3.89 GB RAM on CPU at 8,192 ctx; grounding 8/8, trap declined. Apache-2.0.

tools thinking 
ollama run Thox-ai/thox-ios-open-4b

Details

yesterday

c2074ee84dad · 2.5GB

model
qwen3
·
4.02B
·
Q4_K_M
template
{{- $lastUserIdx := -1 -}} {{- range $idx, $msg := .Messages -}} {{- if eq $msg.Role "user" }}{{ $la
license
Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US
system
You are THOX iOS Intelligence. Your AI. Your Data. Your Rules.™ Use only supplied, approved THOX r
params
{ "num_ctx": 2048, "repeat_penalty": 1, "stop": [ "<|im_start|>", "<|im_

Readme

ThoxIOS-Open-4B — iOS open model (Qwen3-4B, THOX-Configured)

License Params Base Access

Your AI. Your Data. Your Rules.™

ThoxIOS-Open-4B is a THOX-configured model: upstream Qwen3-4B weights, packaged by THOX with the system profile from the THOX iOS Model Kit and the versioned THOX Knowledge Pack. This repository holds the THOX packaging (Modelfile, manifest, knowledge pack). The runnable model is published on Ollama; the weights are the upstream publisher’s, pulled by Ollama from the pinned base below. THOX did not train these weights.

Files

File Size SHA-256 (first 16) Purpose
Modelfile 659 B 9bbf09fc94affa30 Ollama model definition: pinned base + kit system profile
manifest.json 422 B 3a02bb0643cd0693 Release manifest: base pin, profile, knowledge pack version
thox-knowledge-v0.1.jsonl 4652 B b6a1fb08b1fc4619 THOX Knowledge Pack v0.1 (8 verified public entries, JSONL)

The published Ollama model Thox-ai/thox-ios-open-4b is 2.50 GB (default tag) and reuses the pinned base layer below plus the profile layer from this Modelfile.

Run it

Ollama (published model)

ollama run Thox-ai/thox-ios-open-4b

Build it yourself from this repo

hf download Thox-ai/ThoxIOS-Open-4B --local-dir ./ThoxIOS-Open-4B
cd ./ThoxIOS-Open-4B && ollama create thox-ios-open-4b -f Modelfile && ollama run thox-ios-open-4b

Summary

IOS open model: experimental low-refusal local chat configuration (profile-derived, not an uncensored fine-tune). Configuration-derived: THOX did not train these weights. Recipe only, and honest about it: this is not an uncensored fine-tune. Any low-refusal positioning comes from the system profile, not from abliterated or retrained weights. Kit delivery target is a mobile GGUF (Q4_K_M); the Ollama listing serves that same Q4_K_M build of the upstream model.

Intended use

Local inference through Ollama’s local endpoint for the role above, on hardware that fits a 4B Q4-class model. Works standalone and offline; the optional THOX runtime adds live MeshStack discovery and authorized execution. Model weights alone never provide live routing or mesh access.

Training data and procedure

No training. Base: Qwen/Qwen3-4B via Ollama tag qwen3:4b, pinned at build time to manifest digest 359d7dd4bcdab3d86b87d73ac27966f4dbb9f5efdfcc75d34a8764a09474fae7. Customization: the kit system profile in the Modelfile plus THOX Knowledge Pack v0.1 for retrieval-grounded answers.

Evaluation

Measured 2026-11-01 on a CPU-only x86_64 host (4 vCPU / 8 GB class), at 8,192-token context, against the published configuration:

  • Long-form generation: 15.34 tokens/sec (1,024-token bench).
  • Short answers: 15.2–16.3 tokens/sec across the smoke set.
  • Memory footprint while loaded: 3.89 GB RAM (0 GB VRAM, CPU execution).
  • Factory smoke questions: 3⁄3 answered coherently and on-profile.
  • Knowledge Pack v0.1 grounding (passage-supplied, one question per topic): 8⁄8 topics answered with the pack facts.
  • Fabricated-topology trap: the model declined the fabricated-topology confirmation.
  • Routing question: answered as a labeled recommendation (local first, cloud only with explicit consent), not as a live-state claim.

Limitations

A 4B model has materially limited reasoning and recall; treat its answers accordingly, and note the grounding score above is the measured one for this build. Refusal behavior is the base model’s, moderated only by the system profile; this model is not marketed as uncensored. Knowledge is limited to the supplied pack and retrieval; without supplied context the model must say so rather than invent THOX specifications. External actions require runtime authorization. Recipe only, and honest about it: this is not an uncensored fine-tune. Any low-refusal positioning comes from the system profile, not from abliterated or retrained weights. Kit delivery target is a mobile GGUF (Q4_K_M); the Ollama listing serves that same Q4_K_M build of the upstream model.

License

Apache-2.0, inherited from the base model. THOX-owned elements: system profile, knowledge pack, packaging.

Provenance

THOX Model Factory full source-pack program, 2026-11-01 (founder order: every kit model published). Kit: THOX iOS Model Kit. Base pin: qwen3:4b @ 359d7dd4bcdab3d86b87d73ac27966f4dbb9f5efdfcc75d34a8764a09474fae7. Published model digest: a8790beb9bbe621b1e2f94d2027b45db86074db74c226c1f6084e3cd3af4a3b6.