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

HopCoder Mini 35B A3B VL36 Full-SFT

Endpoint-ready full-SFT checkpoint. The training export was repacked into the base model serving key layout with packed MoE expert tensors.

Use Transformers >=5.5.0 for qwen3_5_moe support. vLLM images that ship older Transformers may fail unless configured to use a new enough Transformers backend.

Featherless Compatibility

This repo is prepared for the Featherless Hugging Face ingestion requirements:

• Full model weights, not LoRA or QLoRA adapters.
• Safetensors shards with model.safetensors.index.json.
• Export dtype: FP16.
• Tensor names and shapes are validated against TaimoorSiddiqui/HopCoder-Mini-35B-A3B-VL36; no embedding resize or tensor-shape variation is introduced.
• Model card is present in this repository.

Featherless serves Qwen-family models at up to 16k context. The source config may advertise a larger native context, but prompts plus completions should stay within Featherless' served context limit.

Tool Calling

The tokenizer chat template is included in tokenizer_config.json and chat_template.jinja. Tool calls are trained and documented as JSON inside <tool_call> tags:

<tool_call>
{"name":"tool_name","arguments":{"argument_name":"value"}}
</tool_call>

After a <tool_response> turn, clients should continue generation until the assistant provides a final answer.