Underdog, the on-device assistant developed by Conway Research, released Saluki 27B on October 9, 2026, an Apache 2.0-licensed language model designed to run artificial intelligence agents on personal computers. According to MarkTechPost, it is a 2-bit compressed version of Qwen3.8-27B that takes up 7.89 GB, compared with the 54 GB the full model needs in its original 16-bit format.
The release comes a few days after Underdog presented itself as a privacy-focused alternative to assistants such as Meta’s Muse. Now the company goes a step further and publishes one of the models that can power that kind of assistant without relying on the cloud. The key, according to the company, is that the compression was tuned to protect tool calling, the skill that turns a chat model into an agent able to query services, run functions and chain actions.
To understand the news, it helps to explain quantization. Models store billions of numbers, called weights, that normally take 16 bits each. Reducing that precision to a few bits lets the model fit in far less memory and run on home hardware, but it usually costs quality. The challenge is deciding which parts of the model can be compressed more without losing the capabilities that matter.
Saluki is built in three layers. The base is Qwen3.8-27B, a dense 27-billion-parameter model from the Qwen team, with 64 layers, an architecture that mixes Gated DeltaNet linear attention with gated attention, and a native context of 262,144 tokens. The second layer is a quantized version published by ISTA-DASLab, the group at the Institute of Science and Technology Austria, whose smallest file was 8.4 GB at 2.5 bits per weight.
The third layer is Underdog’s own work, which shrank the file to 7.89 GB and targeted the compression at tool calling. The file is distributed under the name IQ2-mix and, according to MarkTechPost, the company has not published the full recipe for that last pass. The model runs on stock llama.cpp, one of the most widely used tools for running models locally, and on apps built on it; it also includes an optional vision add-on of 629 MB or 928 MB.
The results fall into two groups. In tests run by Underdog in the same harness for both models, Saluki scored 88 on Underdog Bench, a set of 120 tasks drawn from BFCL v4 and frozen before testing, against 84 for the full model and 70 for PrismML’s Bonsai 2. On parallel tool calls, with 100 BFCL v4 tasks and the official checker, Saluki scored 42 against 35 for the original.
Not everything improves. On SWE-bench Verified, with 50 real software issues, Saluki fixed 30 against 33 for the full model. And in comparison with the full model’s public scores, the drop is clear in reasoning and math: 79.2 versus 96.7 on AIME 2025, 80.0 versus 94.6 on AIME 2026, 67.5 versus 79.6 on MuSR and 78.0 versus 83.9 on MBPP+. By contrast, it improves slightly on instruction following, with 93.5 versus 91.5 on IFEval and 72.7 versus 71.0 on IFBench.
Underdog sums up the balance as 96% average retention across nine benchmarks relative to the full Qwen3.8-27B. The best relative result is on parallel tool calls, with 120% retention, and the worst on AIME 2025, at about 82%. MarkTechPost puts it this way: the model beats the 54 GB original at tool calling from a file under 8 GB, but competition math and multi-step reasoning drop by 12 to 18 points.
Comparison with other compressions of the same model helps place it. PrismML’s Bonsai 2 27B is smaller, at 5.95 GB and 1.75 bits per weight, and reports 98.2% retention across 14 benchmarks, with stronger math results, including 95.00 on AIME25. However, it needs PrismML’s own llama.cpp fork, because the stock version rejects its format. Saluki, by contrast, runs on unmodified llama.cpp.
There is an important methodological caveat: each vendor uses its own test harness, so cross-vendor scores are not directly comparable. In addition, Underdog Bench is a test designed by the company itself, although based on tasks from a public set. Until independent evaluators replicate the results, the figures are best treated as developer-reported data.
Why does it matter? AI agents make many calls to the model and, if they run in the cloud, each call has a cost and means sending data to a third party. A 27B-class model that fits in under 8 GB makes it possible to run agents on machines with consumer graphics cards, with full GPU offload, and keep data on the device. For developers and privacy-conscious users, it is a tangible alternative.
The choice to prioritize tool calling over math also reflects a shift in what is expected of small models. For an agent that organizes files, checks a calendar or fills in forms, calling a function correctly matters more than solving an olympiad problem. Underdog is betting that this specialization outweighs the loss in reasoning for most everyday uses.
The release joins a wave of open models competing on efficiency rather than size. The Apache 2.0 license, shared by the base model and the compressions it is compared with, allows commercial use and modification, making it easier for other teams to integrate Saluki into their products or build their own versions from it.
In short, Saluki 27B shows that it is possible to compress a 27-billion-parameter model to less than a seventh of its size while keeping, and even improving, the key skill for agents. The price is a notable drop in math and complex reasoning. For anyone who wants a local agent that uses tools reliably, the offer is attractive; for demanding reasoning tasks, the full model still has the edge.
