TaimoorSiddiqui/HopCoder-Mini-35B-A3B-VL36

HopCoder Mini-35B-A3B-VL36 Overview

HopCoder Mini-35B-A3B-VL36, developed by Taimoor Siddiqui, is a BF16 vision-language model that integrates a 35B-A3B agentic text backbone with a VL36 vision stack. This model is a weight-level merge, preserving the text tensors while adding vision capabilities for image and video understanding. It is specifically designed to support agentic coding and tool-calling workflows, making it suitable for complex development tasks.

Key Capabilities

• Agentic Coding & Tool-Calling: Supports advanced coding and tool-use scenarios.
• Multimodal Input: Processes both image and video inputs via its VL36 vision stack.
• Long-Context Support: Configured for long-context chat and coding applications.
• BF16 Precision: Utilizes BF16 full-precision weights, not a quantized release.

Known Limitations

• Unbenchmarked Vision Quality: The VL36 vision tower's performance has not been formally benchmarked and may differ from natively co-trained VLMs.
• Limited Safety Evaluation: Only basic smoke tests for identity and vision have been conducted; no formal safety or alignment evaluations.
• Agentic Coding Limitations: While capable of reasoning about code, the model struggles with autonomous file editing in unsupervised agent loops, indicating limitations in interactive shell-agent workflows.

Use Cases

This model is particularly well-suited for developers requiring a multimodal LLM for agentic programming tasks, especially those involving code generation, reasoning, and integration with visual data. Its ability to handle long contexts also makes it valuable for detailed coding projects and interactive development environments.