Nanthasit/sakthai-coder-browser

TEXT GENERATIONPricing:Input $0.04 / Cached $0.008 / Output $0.08Concurrent Unit Cost:1Model Size:1.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 30, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Nanthasit/sakthai-coder-browser is a 1.54 billion parameter Qwen2.5-Coder-1.5B-Instruct fine-tune, developed by Nanthasit, designed as a browser automation assistant. It outputs structured XML/JSON tool calls for web interaction, enabling navigation, clicking, typing, and content extraction. The model has a 32,768 token context length and is part of the SakThai Model Family, though its current weights are corrupted and not deployable.

Loading preview...

Model Overview

Nanthasit/sakthai-coder-browser is a 1.54 billion parameter model, fine-tuned from Qwen2.5-Coder-1.5B-Instruct by Nanthasit, specifically for browser automation. It is designed to interact with web pages by generating structured <tool_call> XML/JSON outputs, enabling actions like navigation, clicking elements, typing text, and extracting content.

Key Capabilities (when repaired)

  • Browser Automation: Generates tool calls for web interaction.
  • Structured Output: Produces <tool_call> XML/JSON for actions such as:
    • browser_navigate(url)
    • browser_click(element)
    • browser_type(element, text)
    • browser_extract()
  • Qwen2.5 Tool-Calling Format: Utilizes a custom chat_template.jinja for tool-calling, supporting multi-turn interactions.
  • Context Length: Features a maximum context length of 32,768 tokens.

Current Status and Limitations

CRITICAL: This model is currently BROKEN and NOT DEPLOYABLE. The merged weights are corrupted due to a faulty LoRA merge, specifically affecting attention-projection bias tensors. Evaluation attempts resulted in zero output tokens and whitespace loops, indicating severe functional impairment. Users are advised not to deploy this version until the weights are re-merged and re-verified by the developer. It is text-only, primarily trained on English web interactions, and not servable on HF serverless inference in its current state. A GGUF variant and the unmerged LoRA adapter are also available, but the GGUF variant is also affected by the corrupted weights.