SkillGym-Qwen3.5-9B EXL3 4.00 bpw

Source: https://huggingface.co/reasonwang/SkillGym-Qwen3.5-9B

Source revision: 0b8f92556f6c7ec06bd858327ed2166da8b7697c

ExLlamaV3 revision: 1.5.2+cu128.torch2.10.0

Target bitrate: 4.00 bpw

Architecture: Qwen3_5ForConditionalGeneration

This conversion was produced with ExLlamaV3's official convert.py and passed a local load/generation smoke test before publication.

Local validation

• Generated tokens: 64
• Decode speed: 135.807 tok/s
• GPU: NVIDIA GeForce RTX 4090
• Torch: 2.10.0+cu128

The speed figure is a local smoke benchmark, not a cross-system performance claim.

Compatibility notes

• The source config.json declares one MTP/NextN layer, but the released safetensors do not contain the corresponding mtp.* tensors. This EXL3 conversion therefore omits the nonexistent MTP side model while preserving the released language and vision weights.
• The source repository packages multimodal processor metadata differently from the Qwen3.5 base model; preprocessor_config.json and video_preprocessor_config.json were restored from Qwen/Qwen3.5-9B with matching processor settings for ExLlamaV3 compatibility.
• Local TabbyAPI validation passed text chat, streaming, tool calling, vision, and two concurrent requests.