What is Underdog Saluki 27B 1.0?
Underdog Saluki 27B 1.0 is a 2-bit build of Alibaba's Qwen3.8-27B, published on 7 October 2026 by Conway Research, the a16z-backed company behind the Underdog on-device assistant. It is not a new model: it starts from ISTA-DASLab's 8.4 GB GSQ-RCO quantisation of Qwen3.8-27B and modifies it further, cutting the file to 7.89 GB from about 54 GB at full size. Underdog calls it a preview checkpoint and says the training details will follow.
The single GGUF file runs in stock llama.cpp and Ollama with no custom build. The main file is text only; a separate 629 MB or 928 MB vision file adds image input. Thinking is on by default and can be switched off per request, as with Qwen3.8. Underdog says it runs on a 16 GB laptop.
The weights are under Apache 2.0, the same licence as both models it is built on. In its 30 days to late October it drew about 613,000 Hugging Face downloads, against about 1.27 million for the ISTA build it modifies.
Is Underdog Saluki 27B 1.0 worth it?
Saluki's claim is narrow and useful: at 7.89 GB it keeps most of what Qwen3.8-27B can do as an agent. On Underdog Bench, 120 tool-calling tasks taken from the Berkeley Function Calling Leaderboard and frozen before testing, it passed 88 against 84 for the full 54 GB model and 76 for the ISTA-DASLab build it starts from. It picked the right function in 47 of 48 single-call tasks. On 50 SWE-bench Verified issues it fixed 30 to the full model's 33. Every figure is Underdog's own, run in its own harness, and the company itself says a few tasks of any gap is noise.
The cost of compression lands on reasoning. On AIME 2025 it averages 79.2 against a public 96.7 for the full model, and MuSR falls from 79.6 to 67.5. Parallel calls are shakier than the headline suggests: 42 of 100 passed with strict scoring, and on 23 tasks the reply held no call a checker could read. A lenient parser lifts that to 55. Underdog's own Main build scored 91 on the tool-calling set but browses worse, and the ExLlamaV3 build it calls Mia wins parallel calls with 69, though it is 9.1 GB and will not load in llama.cpp.
What Underdog changed is not documented. The card says only that Saluki is built on the ISTA file and modified; the launch post promises the full training later. For a team deciding between Saluki and the ISTA original, that leaves a 12-task benchmark lead and nothing to reproduce it with.
The practical case is still strong. A standard GGUF that loads in stock llama.cpp and Ollama, fits beside a 16 GB laptop's other work, and keeps Qwen3.8's thinking switch is easier to deploy than any other 2-bit build Underdog compared. The launch post says the release takes text only, while the card, updated a week later, adds a vision file; plan on the vision file as new and lightly documented.
Where it earns its place
A developer building a private desktop assistant on a 16 GB MacBook
Run an agent that calls calendar, file and search tools without sending data out
Serve the 7.89 GB GGUF with llama-server and --jinja, thinking off and temperature 0 for direct calls, as the card recommends. Underdog reports 47 of 48 right-function picks, but budget for a lenient tool-call parser.
An ML engineer comparing 2-bit Qwen3.8-27B builds for an on-device product
Choose between Saluki, the ISTA-DASLab original and ExLlamaV3 builds
Saluki leads on single-tool selection (88 vs 76 for ISTA on Underdog Bench), Mia leads on parallel calls (69 vs 42) but needs ExLlamaV3. Rerun your own tool schemas first, since Underdog's 120-task set is small and the method behind Saluki is unpublished.
A head of IT at a 30-person law firm
Give staff an offline assistant for document chores without a cloud contract
Apache 2.0 allows commercial use with no revenue cap, and the model runs on laptops the firm already owns. It is weak at competition-style maths and letter-level puzzles, which matter little for drafting and lookup.
Before you commit
- Setup
- Download one 7.89 GB file and start llama-server with --jinja -ngl 99 -fa on -c 32768; for Ollama, create a model from a one-line Modelfile. Add --mmproj with the 629 MB or 928 MB vision file for images.
- Integrations
- llama-server exposes the OpenAI chat API on port 8080 with tool calls and a per-request enable_thinking switch; Underdog tested Ollama 0.40.4 on Apple silicon. It also ships inside the Underdog app.
- Data and privacy
- Self-hosted, so data stays on the machine. Apache 2.0 permits commercial use and modification; the NOTICE file credits Qwen and ISTA-DASLab.
- Cost at scale
- Weights are free and there is no first-party API price. The file is 7.89 GB plus up to 928 MB for vision, and Underdog says a 16 GB laptop runs it; no VRAM figure for long contexts is published.
- Limits
- Keeps 82 to 85% of the full model's competition-maths score; weakest on letter-level instruction puzzles; about a fifth of parallel-call replies carry formatting slips; long reasoning with thinking on.
What's good
- 7.89 GB file, from about 54 GB
- 88 of 120 tool-calling tasks vs 84 for full size (self-reported)
- Stock llama.cpp and Ollama, no custom build
- Apache 2.0, like both models it builds on
What's not
- AIME 2025 drops to 79.2 from 96.7
- 23 of 100 parallel-call replies unreadable to a strict checker
- Modification method unpublished; preview checkpoint
- Every benchmark run by Underdog
Best for
Developers running a local tool-calling agent on a 16 GB laptop or a single consumer GPU who accept a maths and reasoning penalty.
Underdog Saluki 27B 1.0 alternatives
Underdog Saluki 27B 1.0: common questions
- Is Underdog Saluki 27B 1.0 worth it?
- The 2-bit Qwen3.8-27B to run when tool calling matters more than maths, as long as you treat Underdog's scores as a promise and not a measurement. We rate it 3.6 out of 5.
- How much does Underdog Saluki 27B 1.0 cost?
- Free Apache 2.0 weights. No hosted API price; you pay for your own hardware, or use it inside the Underdog app.
- What are the downsides of Underdog Saluki 27B 1.0?
- AIME 2025 drops to 79.2 from 96.7; 23 of 100 parallel-call replies unreadable to a strict checker; modification method unpublished; preview checkpoint; every benchmark run by Underdog.
- Who is Underdog Saluki 27B 1.0 best for?
- Developers running a local tool-calling agent on a 16 GB laptop or a single consumer GPU who accept a maths and reasoning penalty.
- How much does Underdog Saluki 27B 1.0 cost for a team or in production?
- Weights are free and there is no first-party API price. The file is 7.89 GB plus up to 928 MB for vision, and Underdog says a 16 GB laptop runs it; no VRAM figure for long contexts is published.
- Is Underdog Saluki 27B 1.0 safe for company data?
- Self-hosted, so data stays on the machine. Apache 2.0 permits commercial use and modification; the NOTICE file credits Qwen and ISTA-DASLab.
- What are the best Underdog Saluki 27B 1.0 alternatives?
- The closest alternatives we have reviewed are Qwen3.8-27B, Gemma 4, gpt-oss-20b and Nemotron 3.5 Bolt (74B-A9B 24GB and E4B 4GB).
Answers checked Oct 28, 2026.