Conway Research Released Underdog Saluki 27B Model
The new 2-bit quantized model retains high-performance tool-calling capabilities in a compact 7.89 GB footprint.
Updated on Oct. 9, 2026 in Artificial Intelligence
Live Poll
Would you use a compressed AI model on your local device for automated tool calling tasks?
Conway Research has released Underdog Saluki 27B, an open-weights model derived from Qwen3.8-27B. The model is available now on Hugging Face under an Apache 2.0 license.
Why it matters
By utilizing 2-bit quantization, this release demonstrates that large language models can maintain complex functional utility even when compressed significantly to fit on consumer hardware. It highlights a growing trend of optimizing high-parameter models for specific tasks like tool calling rather than pure text generation.
Saluki 27B occupies just 7.89 GB, a massive reduction compared to the 54 GB required for the full BF16 Qwen3.8-27B. Despite this compression, it outperformed the base model on Underdog benchmarks with a score of 88, while achieving a 30-issue resolution rate on SWE-bench Verified.
The players
Conway Research
An artificial intelligence research entity focused on developing and releasing optimized, open-source language model architectures.
Hugging Face
A platform and community hub hosting machine learning models, datasets, and infrastructure for the open-source AI ecosystem.
The details
The model achieves this footprint through mixed-precision GGUF quantization, a method that reduces the numerical precision of model weights to save memory. It also integrates Gated DeltaNet linear attention, an architecture that uses gated attention mechanisms to selectively focus on relevant information during inference. This design was specifically tuned to maintain tool calling proficiency despite the aggressive 2-bit compression.
Timeline
October 9, 2026: Underdog Saluki 27B was officially released.
The Tech Race
This model follows a trend of pushing extreme quantization limits to make massive, dense parameter models accessible to local hardware. It marks a departure from standard high-precision deployments by prioritizing functional task execution over raw language density.
Developers can run this model locally using llama.cpp to perform complex tool-calling tasks without requiring enterprise-grade GPU clusters. Users also have the option to integrate 629 MB or 928 MB vision add-ons depending on their specific multimodal requirements.
The takeaway
Saluki 27B proves that intentional quantization can maintain, and even improve, task-specific performance compared to uncompressed models. Watch for community-driven benchmarks on standard platforms to confirm if this 2-bit efficiency holds up across broader generative reasoning tasks.
Further reading
Explore more developments in the Artificial Intelligence sector on our dashboard.
Live Poll
Would you use a compressed AI model on your local device for automated tool calling tasks?
