Thox-ai

ThoxEdge-RustCoder-0.5B

README

License: apache-2.0

Summary

ThoxEdge-RustCoder-0.5B is a Rust-leaning fine-tune trained by THOX.ai on a public Apache-2.0 Rust code corpus (74,606 filtered code-producing examples): a LoRA fine-tune (rank 32), merged into the base weights. It serves cheap Rust completion in ThoxRoute's lowest tier. It is a 0.5B model and is not a strong Rust coder; the numbers below are a same-harness delta measurement, nothing more. It is a hosted/desktop model: its GGUF ladder spans only about 260 MB because a large share of its parameters sit in an embedding that does not quantize down, so it does not fit THOX's smallest devices.

Highlights

THOX fine-tune: a LoRA fine-tune (rank 32), merged into the base weights.
This repo ships safetensors only (no GGUF): bash hf download Thox-ai/ThoxEdge-RustCoder-0.5B --local-dir ./ThoxEdge-RustCoder-0.5B

Specifications

Table

Base model	Qwen/Qwen2.5-Coder-0.5B-Instruct
Parameters	0.5B
Formats	Safetensors weights
Training	LoRA fine-tune (rank 32), merged into the base weights
License	Apache-2.0. The THOX fine-tune and additions in this repo are covered under Apache-2.0; the training corpus is also Apache-2.0. See the license text for terms.
This repository contains	Full weights and tokenizer in this repo

Files

Table with columns: File, Size, Purpose
File	Size	Purpose
`model.safetensors`	988,097,824 B	Model weights (safetensors)
`tokenizer.json`	11,421,892 B	Tokenizer files
`chat_template.jinja`	2,507 B	Chat template
`.gitattributes`	1,570 B	Repository file
`config.json`	1,284 B	Model configuration
`tokenizer_config.json`	693 B	Tokenizer files
`generation_config.json`	217 B	Generation configuration

Run it

This repo ships safetensors only (no GGUF):

bash

hf download Thox-ai/ThoxEdge-RustCoder-0.5B --local-dir ./ThoxEdge-RustCoder-0.5B

python

from transformers import AutoModelForCausalLM, AutoTokenizertok = AutoTokenizer.from_pretrained("Thox-ai/ThoxEdge-RustCoder-0.5B")model = AutoModelForCausalLM.from_pretrained("Thox-ai/ThoxEdge-RustCoder-0.5B")

Prompting matters. Ask for a complete function body in plain terms, for example "Write the complete body of the Rust function below." Do not prompt with "Complete the Rust code. Output only the missing code." — that phrasing matches training rows whose targets are short fragments, and this checkpoint answers it with a fragment (measured 2.56% where the recommended wording scores 28.21%).

Platforms

Table with columns: Platform / device, Run path
Platform / device	Run path
Hugging Face	Download the builds listed under Files from this repo (see Run it)

Intended use

Cheap Rust code completion on hosts, inside the THOX stack.
Out of scope: medical, legal, financial, safety-critical, or other high-stakes decisions without human review; unattended control of physical devices without a confirmation step.

Training data and procedure

THOX fine-tune of Qwen/Qwen2.5-Coder-0.5B-Instruct. The corpus and procedure are described in this card as published at release (see Summary and Evaluation); no training beyond what this card describes is claimed for this repository.

Evaluation

MultiPL-E HumanEval-Rust, 156 problems, executed (each generation is compiled and its assertions run), greedy decoding, both models scored in the same job:

Table with columns: Model, Completion mode, Chat mode
Model	Completion mode	Chat mode
Base model	21.15% (33/156)	25.64% (40/156)
ThoxEdge-RustCoder-0.5B	25.64% (40/156)	28.21% (44/156)
Delta	+4.49 pp	+2.57 pp

The noise floor of this harness is about ±1 problem (±0.64 pp); both deltas clear it, and anything under about 1.5 pp should be treated as a tie. Assembly is deliberately generous (a body-only or whole-function answer passes if either works), so these numbers are not comparable to published MultiPL-E leaderboards; the delta between the rows is the measurement. No THOX GATE49 result is published for this model.

Limitations

It fails roughly seven of every ten HumanEval-Rust problems (28.21% pass@1 in its best mode).
A training-renderer defect placed auxiliary output fields inside the code fence in about half the rows of three training categories; it caused no measured regression, and this checkpoint retains the prompt sensitivity described above.
Not sized for THOX's smallest devices (see summary).

Provenance

Published by THOX.ai as Thox-ai/ThoxEdge-RustCoder-0.5B. Upstream base: Qwen/Qwen2.5-Coder-0.5B-Instruct. This card was conformed to THOX Model Card Standard v1.0 on 2026-10-31 (frontmatter, tags, header, files, and section structure); the technical content is as published at release.

License

Apache-2.0. The THOX fine-tune and additions in this repo are covered under Apache-2.0; the training corpus is also Apache-2.0. See the license text for terms.

THOX.ai — Your AI. Your Data. Your Rules.™

Citation

bibtex

@misc{thoxai2026thoxedgerustcoder05b,      title={ThoxEdge-RustCoder-0.5B},      author={THOX.ai},      year={2026},      url={https://huggingface.co/Thox-ai/ThoxEdge-RustCoder-0.5B}}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Thox-ai

Model Tree

Base

Qwen/Qwen2.5-Coder-0.5B-Instruct

Fine-tuned

this model

Input Modalities

Text

Output Modalities