groxaxo/Qwento-Agentic
Qwento-Agentic by groxaxo is a 35.1 billion parameter model checkpoint, fine-tuned from Qwen/Qwen-AgentWorld-35B-A3B, a Qwen3.5 MoE architecture. This experimental model, utilizing QLoRA, is specifically designed and trained on curated datasets for coding tasks, including reasoning, trace-inversion problem solving, and multi-turn agentic coding. It is intended for open-source evaluation and reproducible experimentation in coding-focused applications.
Loading preview...
Overview
Qwento-Agentic is an experimental model checkpoint published by groxaxo, built upon the Qwen/Qwen-AgentWorld-35B-A3B base model. This 35.1 billion parameter model (with approximately 3 billion active parameters due to its MoE architecture) is a QLoRA fine-tune, merged into BF16, and is specifically designed for coding tasks.
Key Capabilities & Features
- Base Architecture: Derived from Qwen3.5 MoE, featuring a hybrid DeltaNet linear-attention and full-attention mechanism with 256 experts.
- Fine-tuning Method: QLoRA (rank 16, α 32) applied to the model's sequence-mixing path, with MoE experts frozen.
- Training Data: Trained on a curated blend of publicly available datasets focusing on:
- Reasoning and trace-inversion problem solving (
Jackrong/Claude-opus-4.7-TraceInversion-5000x,lordx64/reasoning-distill-claude-opus-4-7-max,lordx64/reasoning-distill-opus-4-7-max-sft) - GPU-kernel coding traces (
Infatoshi/kernelbench-mega-traces) - Multi-turn agentic coding and tool use (
Glint-Research/fable-5-traces)
- Reasoning and trace-inversion problem solving (
- Format: BF16 safetensors, compatible with Hugging Face Transformers and vLLM.
Limitations
As a preliminary test checkpoint from a short training run, Qwento-Agentic has not been benchmarked and should be treated as experimental. It inherits the license and any usage restrictions of its base model, Qwen/Qwen-AgentWorld-35B-A3B.