Thox-ai/ thoxroute

3 1 week ago

THOX.ai model-routing classifier. Clean-room Qwen2.5-1.5B fine-tune that picks the best THOX model for a query. Infrastructure.

1.5b 
ollama run Thox-ai/thoxroute:1.5b

Models

View all →

1 model

thoxroute:1.5b

986MB · 32K context window · Text · 1 week ago

Readme

ThoxRoute-1.5B

THOX.ai’s model-routing classifier: given a user query and a set of candidate routes, it returns the best route for the THOX model stack.

Tags: router classifier qwen2 1.5b q4_k_m infrastructure

At a glance

  • Base model: Qwen/Qwen2.5-1.5B-Instruct (Apache-2.0)

  • - Fine-tune: LoRA SFT on THOX’s own routing taxonomy (clean-room)

  • - Architecture: qwen2

  • - Parameters: 1.5B

  • - Quantization: Q4_K_M (986 MB)

  • - Native context: 32768 tokens

  • - Prompt template: ChatML; the system prompt lists candidate routes, the model replies with one route name

  • - License: Qwen2.5-1.5B-Instruct base: Apache-2.0. THOX fine-tune, Modelfile, and packaging: © 2026 THOX.ai LLC.

    • Publisher: THOX.AI LLC

      Why clean-room

THOX evaluated customizing or distilling katanemo/Arch-Router-1.5B. That model is under the Katanemo Community License (DigitalOcean): commercial use is prohibited without a separate licence, and any model trained on it or its outputs must be named beginning with “DigitalOcean”. ThoxRoute is therefore built from scratch on Apache-2.0 Qwen2.5-1.5B using THOX’s own route taxonomy. No Arch-Router weights or outputs are used at any stage.

Quick start

ollama pull thox-ai/thoxroute:1.5b
ollama run thox-ai/thoxroute:1.5b "my rust build fails with a borrow checker error"
# -> thoxforge-7b

The routing instructions and route list ship in the model’s SYSTEM prompt, so it routes out of the box. Run greedily (temperature 0, already the default).

Results (v0.2)

Eval Result
Held-out validation (289) 289⁄289 = 100%
Fresh out-of-distribution probe (28 hand-written) 26⁄28 = 92.9%
Invalid / hallucinated route names 0

The 100% is not the headline: that split shares templates with the training data. The out-of-distribution probe is the honest signal, and it is deliberately hostile - its non-English items use languages that appear nowhere in training (Dutch, Polish, Turkish, Swedish, Greek, Indonesian, Czech), of which it routed 7⁄8 correctly. The model never emitted an invalid route name.

Intended use

Infrastructure component, not a chat assistant. It selects which THOX model should serve a query (consumed by thoxcore-router and the router-shim).

Not intended for

General chat, reasoning, or code generation - use the THOX assistant models for that.

Notes

  • Routing: express routes by positive signal, not negation (“NOT ”).

  • - Known gaps: a non-English coding question may route by task rather than language; some conversationally-phrased creative requests fall to other.

    • On-device: built for local execution on THOX hardware and desktops. Your AI. Your Data. Your Rules.

      Links

  • THOX: https://www.thox.ai

    • Press: press@thox.ai Published by THOX.AI LLC from the thoxroute-build pipeline.