josephmayo/von3b

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:3.1BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 23, 2026Architecture:Transformer Featherless Exclusive Cold

josephmayo/von3b is a 3.1 billion parameter model designed for offline coding assistance and autonomous agent capabilities. Developed by josephmayo, it was fine-tuned from WeiboAI/VibeThinker-3B using SFT, RL, and group-conditioned adaptive LoPD. This model excels at code generation and repair, and uniquely emits one-line tool calls for agentic functions, running locally on standard 8GB machines without network access.

Loading preview...

Overview

VON-3B is a 3.1 billion parameter model developed by josephmayo, specifically engineered to function as both an offline coding assistant and an autonomous agent. It is designed to run locally on standard machines with 8 GB RAM, such as laptops, without requiring an API key or network connection during inference. The model was fine-tuned from WeiboAI/VibeThinker-3B using a comprehensive pipeline including SFT, RL, group-conditioned adaptive LoPD, LoRA, and weight edits.

Key Capabilities

  • Code Generation and Repair: Optimized to write and fix code efficiently.
  • Autonomous Agent Functionality: Capable of emitting real one-line tool calls (<tool_call>{...}</tool_call>), enabling it to act as an agent rather than just a chatbot.
  • Concise Reasoning: Designed to keep reasoning short, improving inference speed.
  • Offline Operation: Runs entirely locally, making it suitable for environments without internet access.

Performance Highlights

On a matched EvalPlus 0.3.1 HumanEval check (164 tasks), VON-3B significantly outperforms its base model, VibeThinker-3B:

  • HumanEval pass@1: 0.921 (151 / 164) compared to 0.866 (142 / 164) for the base.
  • HumanEval+ pass@1: 0.884 (145 / 164) compared to 0.817 (134 / 164) for the base.

Furthermore, in a tool probe of 32 tasks, VON-3B successfully emits a valid one-line <tool_call> with short reasoning on 32 out of 32 tasks, whereas the base model emits 0 out of 32. The model supports a context length of 32,768 tokens, allowing for long coding sessions.

Good For

  • Developers needing a local, offline coding assistant.
  • Applications requiring autonomous agent capabilities with tool-use.
  • Resource-constrained environments (e.g., laptops with 8GB RAM) where network access is limited or unavailable.

Finetunes
1 models