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

Hugging Face
TEXT GENERATIONConcurrent Unit Cost:3Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 5, 2026License:otherArchitecture:Transformer Featherless Exclusive Warm

The TaimoorSiddiqui/HopCoder-Mini-35B-A3B-VL36-fullsft is an endpoint-ready, full-SFT checkpoint based on the Qwen-family architecture, featuring packed MoE expert tensors. This model is specifically trained for tool calling, with a defined JSON format for tool interactions. It is optimized for deployment environments like Featherless, supporting up to 16k context length for prompts and completions.

Loading preview...

Model Overview

The TaimoorSiddiqui/HopCoder-Mini-35B-A3B-VL36-fullsft is a fully instruction-tuned (SFT) checkpoint, designed for immediate deployment. It utilizes a Qwen-family architecture with packed Mixture-of-Experts (MoE) tensors, making it efficient for serving. Users should ensure they are running Transformers version >=5.5.0 for proper qwen3_5_moe support.

Key Capabilities

  • Tool Calling: The model is explicitly trained for tool calling, using a structured JSON format within <tool_call> tags. This enables seamless integration with external functions and APIs.
  • Deployment Ready: Provided as a full model weight (not LoRA or QLoRA), with safetensors shards and model.safetensors.index.json, making it compatible with platforms like Featherless.
  • FP16 Export: The model weights are exported in FP16 dtype for optimized performance and reduced memory footprint.
  • Context Length: While the native configuration may support larger contexts, it is optimized for up to 16k context length for prompts and completions in deployment environments like Featherless.

Usage Notes

After a <tool_response> turn, clients are expected to continue generation until the model provides a final answer, indicating the completion of the tool-assisted task.