⚡️ You don't need Unsloth's "good enough" Qwen3.8-27B quant anymore. BaseQuant… · @QwenAIFans · TwKit

@QwenAIFans

⚡️ You don't need Unsloth's "good enough" Qwen3.8-27B quant anymore. BaseQuant ships the same file sizes — and stays ~15% closer to full BF16 on agent + tool prompts (KL divergence). Medium: 16.41 GB ≈ UD-Q4_K_M Small: 10.93 GB ≈ UD-IQ3_XXS Agents call the model over and over. Tiny drift compounds. That's why KL matters more than one-shot accuracy. Caveat: this is distribution closeness, not a new SWE-bench score — and on Large the public-prompt gap shrinks a lot. via Hugging Face Blog 👇 https://huggingface.co/blog/basecompute/base-quant-v1

曝光 161 · 评论 0 · 点赞 1 · 书签 0