Reportedly: Underdog AI releases Saluki 27B: a 2-bit quantized Qwen3.8-27B at 7.89 GB
Model: Qwen3.8-27B · availability, license and releases
Reported by IT Home · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Underdog AI released Saluki 27B, a 2-bit quantized version of Qwen3.8-27B with a file size of 7.89 GB, IT Home reported on October 10, 2026. The model is designed for agent applications and excels at tool use, outperforming the full-size Qwen3.8-27B in tool selection and parallel tool calling. It uses standard llama.cpp format and loads without custom compilation.
China context
- Outside China
- Not stated in the sources yet
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Builders can integrate Saluki 27B into llama.cpp-based applications without custom compilation, enabling local agent development with a 7.89 GB model.
- For investors
- Investors should monitor whether Underdog AI's focus on efficient agent models gains traction among developers seeking cost-effective local inference.
Translated from Chinese. Quotes and facts link to the original sources.
Underdog AI has released Saluki 27B, a 2-bit quantized and fine-tuned version of Qwen3.8-27B. The model retains only text input/output capabilities and has a file size of 7.89 GB. According to Underdog, Saluki is designed for agent applications, with strengths in everyday reasoning and tool use, while its math performance is relatively weaker. Its tool selection and multi-tool parallel calling performance even surpasses the full-size Qwen3.8-27B. Saluki uses the standard llama.cpp format and can be loaded in standard llama.cpp and applications built on it without any custom compilation.
Saluki 27B is a 2-bit quantized version of Qwen3.8-27B, reducing the model size to 7.89 GB. It is optimized for agentic tasks, showing improved tool selection and parallel tool calling compared to the full-size model, though math performance is weaker. The model is distributed in llama.cpp format, enabling easy integration with existing llama.cpp-based applications.
Developers outside China can now run a 27B-parameter agent-focused model on consumer hardware due to its 7.89 GB size, lowering the barrier to local agent deployment. This directly challenges larger models that require more expensive infrastructure for similar tool-use capabilities.
The 7.89 GB size enables deployment on edge devices and local machines, reducing cloud inference costs for agent applications. The focus on tool use may attract developers building autonomous agents who need efficient, local models.
The next observable signal is whether Underdog AI releases benchmarks or independent evaluations confirming the claimed tool-use improvements. Another signal is adoption in open-source agent frameworks that rely on llama.cpp.