LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
← Voltar para a leitura
NotíciasTópico · Saluki 27B模型Visão rápida1 min de leituraEditorial DisconfirmAI

Underdog Releases Saluki 27B: 7.89GB Quantized Model Self-Reports Better Tool Calling Than Original

Underdog released Saluki 27B, a 2-bit quantized model based on Qwen3.8-27B weighing just 7.89GB, which self-reports superior tool-calling performance compared to the 54GB original.

Verificado 2026-10-09 17:28 GMT+8

Underdog (Conway Research) has released Saluki 27B, a 2-bit quantized GGUF model based on Qwen3.8-27B with a file size of only 7.89GB.

The model targets local developers seeking agent workflow support on standard llama.cpp runtimes. According to the official model card, Saluki scored 88 on "Underdog Bench" (120 tasks from BFCL v4), surpassing the 54GB full-precision original's score of 84. In parallel tool-calling tests, Saluki achieved 42 points versus the original's 35.

However, this compression involves significant trade-offs. Official data shows Saluki's performance on competition math (AIME 2025) dropped to 79.2 from the original's 96.7, retaining about 82% of capability. Additionally, some benchmark comparisons rely on public scores from different testing frameworks, and all results are vendor self-reported without independent reproduction so far.

Fontes

  • huggingface.co↗
Tópico

Saluki 27B模型

Underdog发布基于Qwen3.8-27B的2-bit量化模型Saluki,体积仅7.89GB。该模型在工具调用任务上自报优于54GB的原版模型,但在数学推理和复杂编码任务中性能显著下降。

Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

2 etapas1 peçasAtualizado 2026-10-09 18:18

Como chegamos aqui

  1. 2026-10-08Underdog发布Saluki 27B,基于Qwen3.8-27B的2-bit GGUF格式,大小7.89GB,支持stock llama.cpp。huggingface.co ↗
  2. 2026-10-09MarkTechPost等媒体报道称,Saluki在工具调用基准测试中得分88,超过全精度模型的84分,但数学推理得分大幅下降。marktechpost.com ↗

O que ainda está indefinido

  • Saluki的具体量化训练配方(IQ2-mix pass)尚未公开,社区能否独立复现其性能提升?
  • 在实际代理工作流中,因量化导致的格式错误(约20%的并行调用回复有小格式瑕疵)对系统稳定性的影响有多大?
  • 与PrismML Bonsai 2等其他小型化模型相比,Saluki在真实场景下的性价比和兼容性优势是否持续存在?
Ler o tópico completo →
Continuar pesquisando →