Huihui Qwen3.8 Flash Next Gets Swift GGUFs

This is a meaningful inference experiment, not just another abliterated model mirror—but the repository calls it a test/validation update, so treat performance gains as promising rather than settled.

• –Swift 1.5 claims 63.4% fewer thinking tokens and 1.8× faster inference with under 1% accuracy loss versus its base model. [Swift model card](https://huggingface.co/ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF)
• –Available IQ2_XS and IQ3_XXS builds are roughly 68GB and 76GB, making this 177B-parameter model more approachable for high-memory local systems.
• –Current llama.cpp support and Strata compatibility give developers practical OpenAI-compatible local serving options.
• –The abliterated variant reduces refusal behavior, which expands experimentation but raises safety and deployment concerns.
• –The key question is whether Swift’s token savings hold across real coding and agent workloads, not only model-card evaluations.