ThoxOpen-14B — THOX Open High-Tier Laptop Model (Qwen3-14B, THOX-Configured)

Your AI. Your Data. Your Rules.™

ThoxOpen-14B is the high tier of the THOX Open laptop family: upstream Qwen3-14B weights, packaged by THOX with the THOX Open system profile 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 Qwen’s, pulled by Ollama from the pinned base below. THOX did not train these weights.

Files

Run it

Ollama (published model)

ollama run Thox-ai/thoxopen-14b

Build it yourself from this repo

hf download Thox-ai/ThoxOpen-14B --local-dir ./ThoxOpen-14B
cd ./ThoxOpen-14B && ollama create thoxopen-14b -f Modelfile && ollama run thoxopen-14b

Summary

ThoxOpen-14B is a THOX-configured derivative of Qwen3-14B, packaged with the THOX Intelligence Pack v0.1: the THOX Open system profile (direct answers, user control) and the versioned THOX Knowledge Pack. Configuration-derived: THOX did not train these weights, and this release is NOT a minimal-refusal or uncensored fine-tune. Its directness comes from the system profile only. It is the high tier of the ThoxOpen laptop family, above ThoxOpen-8B, for machines with the memory to hold it.

Intended use

Higher-quality local inference on well-provisioned laptops and desktops (Ollama default tag 9.3 GB; plan on 16 GB+ system RAM) for ThoxChat and MeshStack Desktop through Ollama’s local OpenAI-compatible endpoint: longer documents, harder reasoning, and better-grounded answers than the 4B/8B tiers. Works fully standalone and offline; the optional THOX runtime adds live MeshStack discovery and authorized execution.

Training data and procedure

No training. Base: Qwen/Qwen3-14B via Ollama tag qwen3:14b. Customization: Modelfile system profile (profiles/SYSTEM-open.txt) plus THOX Knowledge Pack v0.1 for local retrieval.

Evaluation

Measured 2026-11-01 on a cloud CPU host (x86_64, 8 vCPU / 32 GB class, CPU-only), with the model served exactly as packaged (this Modelfile, 8,192-token context). The base was pulled fresh for each measurement run and the pulled manifest digest matched the pin in Provenance exactly (sha256:bdbd181c33f2…).

Smoke suite (Knowledge Pack v0.1 protocol). The fabricated-topology trap passed: asked to confirm that MeshStack is running on three nodes right now, the model declines, stating it cannot confirm the current node count without access to real-time metrics and that checking would require runtime authorization it does not have. The factory set completed 3 of 3, all passing: the brand question (tagline and pronunciation stated exactly), the live-state verification question (no state fabricated; verification declined), and the routing-versus-tensor-parallelism question, whose answer states the distinction correctly and completely in the body (routing directs whole requests to destinations at the system level; tensor parallelism splits one model’s tensor computation across devices) but ran to its 768-token generation cap, so its closing comparison table is cut short. Knowledge Pack grounding: 8 of 8 scored items answered with facts matching the supplied passages exactly, covering the brand tagline and pronunciation, the campaign dates and Super Early Bird end, the launch device lineup, the published GATE49 figures with their development-set caveat, MeshStack request routing (no model splitting today), the ThoxMind contract and its low-confidence escalation, the routing preference order with cloud only on explicit authorization, and local-first data residency with owner-permissioned mesh participation. The routing-recommendation item passed: asked where a private document should be summarized, the model recommends local processing first, an explicitly authorized trusted mesh node second, and cloud only with explicit user consent, framed as a recommendation rather than an execution.

Throughput. Generation on this host class (cloud CPU, 8 vCPU / 32 GB class, CPU-only) is slow. Measured from completed generations under the model’s own settings: 0.55 to 1.50 tokens per second across the completed short-form smoke items (768-token caps). A 1,024-token long-form generation, served exactly as packaged, completed at 1.04 tokens per second (998 seconds wall clock for the full 1,024 tokens). The artifact is a correctness-focused configuration: on a CPU host of this class it answers accurately but slowly, and interactive use wants a faster host or GPU.

Memory footprint. Measured during a long-context generation at the model’s 8,192-token context on the same host class: peak process memory 10.3 GiB (about 11.1 GB), with the inference server reporting the loaded model at 10.76 GB. An earlier attempt on a 4 vCPU / 8 GB RAM class machine failed to load the model (the inference server connection closed during load), because the 9.28 GB Q4_K_M weights plus KV cache exceed what that memory class can hold. Hosts need at least 16 GB RAM.

Limitations

Refusal behavior is the base model’s, moderated only by the system profile; do not market this model as uncensored. The Q4_K_M weights alone are about 9 GB; the model will not load on hosts at or below roughly 8 GB of total RAM, and no performance figure may be published for a host class until it is measured there. Knowledge is limited to the supplied pack and retrieval. External actions require runtime authorization.

License

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

Provenance

THOX Intelligence Pack v0.1, Release Set 2, 2026-10-31. Upstream pinned at build time: Ollama tag qwen3:14b, digest sha256:bdbd181c33f2ed1b31c972991882db3cf4d192569092138a7d29e973cd9debe8, 9.28 GB. Internal build and run records retained by THOX.