groxaxo/Qwento-Agentic

Qwento-Agentic

<!-- polished-overview:start -->

Overview

Qwento-Agentic is a model checkpoint packaged for compatible Hugging Face runtimes, published by `groxaxo`. It is intended for open-source evaluation, reproducible experimentation, and compatible local or hosted inference workflows. The wording below is deliberately limited to what can be verified from this repository's metadata and artifacts.

At a glance

What is included

• —*.safetensors (21 files)
• —config.json
• —generation_config.json
• —tokenizer.json
• —tokenizer_config.json
• —chat_template.jinja
• —Additional configuration, tokenizer, processor, or shard files (33 visible artifacts total)

Quick start

Getting started

Start with the upstream library named in the repository metadata and keep all configuration, tokenizer, processor, and weight files together. This repository is an artifact release, so the source project remains the authoritative reference for task-specific loading code.

Compatibility and responsible use

• —Use a runtime that explicitly supports this format, architecture, and modality.
• —Keep configuration, tokenizer, processor, projection, and weight files from the same revision together.
• —Review the source model card and license before redistribution or deployment.
• —Hardware needs depend on parameter count, context length, cache precision, quantization, and concurrency.
• —Report reproducible issues with the runtime version, hardware, launch command, and a minimal example.

Generated outputs may be inaccurate or unsuitable for a given use case. Users are responsible for testing behavior, applying appropriate safeguards, and complying with applicable licenses and laws. <!-- polished-overview:end -->

⚠️ Test run. This is an early experimental checkpoint, not a finished model.

A QLoRA fine-tune merged into BF16, built on top of `Qwen/Qwen-AgentWorld-35B-A3B` (a Qwen3.5 MoE: 35B total / ~3B active, hybrid DeltaNet linear-attention + full-attention, 256 experts). It was trained on a curated set of publicly available datasets and is designed for coding tasks.

What this is

• —Type: test run — a single short curriculum stage (2K sequence length), early checkpoint.
• —Method: QLoRA (rank 16, α 32) applied to the model's sequence-mixing path (full-attention q/k/v/o + linear-attention input/output projections across all 40 layers), then merged into the BF16 base weights. The 256 MoE experts were left frozen.
• —Format: BF16 safetensors, drop-in with 🤗 Transformers / vLLM (same architecture and tokenizer as the base).

Training data (curated, publicly available)

A token-balanced blend of cleaned public datasets:

All sources were structurally cleaned and quality-filtered before mixing.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained(
    "groxaxo/Qwento-Agentic", torch_dtype="bfloat16", device_map="auto", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("groxaxo/Qwento-Agentic")

Limitations

This is a preliminary test checkpoint from a short training run; it has not been benchmarked and should be treated as experimental. It inherits the license and any usage restrictions of the base model (Qwen/Qwen-AgentWorld-35B-A3B).