Saluki 27B
An open-weight model with 27 billion parameters released by Conway Research. It is compressed to approximately 7.89 GB using IQ2 quantization, features strong tool-calling capabilities, is licensed under Apache 2.0, and supports local deployment.
Timeline (last 90 days)
As performance data published from a single source, the above benchmark scores still require broader community replication and verification, especially the counter-intuitive conclusion that tool-calling performance surpasses the original full-size model.
The Underdog AI team released and fully open-sourced the on-device large model Saluki 27B 1.0.
As measured by official testing across nine comprehensive authoritative benchmarks, Saluki retains an average of up to 96% of the original model's performance.
In 120 function-calling tests on Underdog Bench 120—built on Berkeley's BFCL V4 leaderboard—Saluki 27B scored 88, surpassing the full-size original model's score of 84.
In the BFCL official 100-item parallel tool-calling test, Saluki scored 42, surpassing the original model's 35, representing approximately a 20% improvement in success rate.
In 50 real issues from the SWE-Bench Verified benchmark, Saluki successfully resolved 30, while the original model resolved 33, retaining approximately 91% of real-world bug-fixing capability.
Saluki scored 93.5 on the IFEval benchmark (versus the original model's 91.5), and scored 72.7 on IFBench, surpassing the original model's 71.0.
Saluki scored 78.0 on MBPP+, 67.5 on multi-step causal reasoning, and 80.0 on AIME, retaining 82%–85% of the original model's mathematical reasoning capability.
The Saluki main model focuses on text and code logic, supports a 32K context, and requires an additional download of approximately 0.9 GB of the visual projection plugin M-Projector when image processing is needed.
Saluki uses the standard GGUF format, can be launched with a single unmodified llama.cpp command, and natively supports the Qwen3 conversation template.
12 more timeline events
All Facts (20)
Saluki uses the standard GGUF format, can be launched with a single unmodified llama.cpp command, and natively supports the Qwen3 conversation template.
50%UnverifiedAs performance data published from a single source, the above benchmark scores still require broader community replication and verification, especially the counter-intuitive conclusion that tool-calling performance surpasses the original full-size model.
50%UnverifiedSaluki supports flexible switching of thinking modes, with long-chain-of-thought enabled by default, and the ability to disable it on a per-request basis during high-frequency tool calls for fast response.
50%UnverifiedAs measured by official testing across nine comprehensive authoritative benchmarks, Saluki retains an average of up to 96% of the original model's performance.
50%UnverifiedIn 120 function-calling tests on Underdog Bench 120—built on Berkeley's BFCL V4 leaderboard—Saluki 27B scored 88, surpassing the full-size original model's score of 84.
50%UnverifiedIn the BFCL official 100-item parallel tool-calling test, Saluki scored 42, surpassing the original model's 35, representing approximately a 20% improvement in success rate.
50%UnverifiedIn 50 real issues from the SWE-Bench Verified benchmark, Saluki successfully resolved 30, while the original model resolved 33, retaining approximately 91% of real-world bug-fixing capability.
50%UnverifiedSaluki scored 93.5 on the IFEval benchmark (versus the original model's 91.5), and scored 72.7 on IFBench, surpassing the original model's 71.0.
50%UnverifiedSaluki scored 78.0 on MBPP+, 67.5 on multi-step causal reasoning, and 80.0 on AIME, retaining 82%–85% of the original model's mathematical reasoning capability.
50%UnverifiedThe Saluki main model focuses on text and code logic, supports a 32K context, and requires an additional download of approximately 0.9 GB of the visual projection plugin M-Projector when image processing is needed.
50%UnverifiedSaluki 27B uses the Apache 2.0 open-source license, allowing free commercial use and derivative development
50%UnverifiedSaluki 27B shows a regression compared to the original in competitive mathematics and complex reasoning tasks
50%UnverifiedIf the application scenario revolves around tool calling, agent orchestration, or API integration, Saluki 27B is a cost-effective choice; if the core requirement is mathematical reasoning or complex inference, the full-precision original version is more reliable
50%UnverifiedUnderdog Saluki 27B is a 2-bit GGUF quantized version built on Qwen3.8-27B
50%UnverifiedSaluki 27B has a model size of 7.89 GB, while the original model size is 54 GB
50%UnverifiedSaluki 27B is compressed by nearly 85% in size compared to the original
50%UnverifiedSaluki 27B surpasses the full-precision original model in tool-calling capability
50%UnverifiedSaluki 27B scored 88 on tool calling and is licensed under Apache 2.0.
50%UnverifiedSaluki 27B can run locally on CPU or GPU via llama.cpp and also supports Docker containerized deployment.
50%UnverifiedBonsai将27B模型从全精度的54GB压缩至3.9GB
50%