Most community Qwen forks reproduce a base and add weights; this release instead focuses on low-level numerical repair and packaging for GGUF runtimes. The maintainer scanned and corrected a small number of mis-scaled tensors and applied a Sig-ScaleSync-Genesis-SVD procedure so the model behaves like a healthy 35B MoE Qwen3.6 in GGUF form while including Hermes-agent tuning and multimodal support.
Key Capabilities
- Tensor-level repair (Sig-ScaleSync-Genesis-SVD): fixes scale mismatches and noisy blocks in specific ssm_conv1d and attention tensors, so what: reduces distribution drift and saturation errors without retraining, improving numerical stability for inference and quantization workflows.
- Hermes agent packaging and function-calling dataset prep: includes agent-oriented fine-tuning artifacts and compatibility with NousResearch/hermes-function-calling-v1, so what: makes the model more immediately useful for agentic workflows and multimodal prompt patterns.
- GGUF + runtime compatibility: packaged for GGUF, tested recommendations for APEX or Q8_K_P quant, and works with llama.cpp/LM Studio/koboldcpp, so what: lowers friction for local deployment on common inference stacks.
- Multimodal and MoE architecture preserved: supports image-text-to-text pipelines and uses a 35B MoE design, so what: retains large-context and multimodal capabilities while enabling lower active-parameter costs per forward pass.
Who it's for & tradeoffs
Great fit if you run local GGUF inference or experiment with agentic multimodal LLMs and need a repaired 35B Qwen variant that tolerates quantization and long-context workloads. Look elsewhere if you require an officially maintained release from the original Qwen vendor, need formal licensing assurances beyond the packaged apache-2.0 tag, or you cannot accept an 'uncensored' base model. Expect to validate safety and behaviour for production use and to follow recommended quant/advice for specific GPUs.
