# Underdog Saluki 27B

> Source: https://aiwiki.ai/wiki/underdog_saluki_27b
> Updated: 2026-10-15
> Fact-checked: 2026-10-15
> Categories: AI Inference, AI Models, Large Language Models, Open Source AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Underdog Saluki 27B." aiwiki.ai, 15 Oct 2026. https://aiwiki.ai/wiki/underdog_saluki_27b
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Underdog Saluki 27B** is a 2-bit [quantized](https://aiwiki.ai/wiki/quantization) build of Alibaba's open-weight [Qwen3.8](https://aiwiki.ai/wiki/qwen3_8)-27B language model, released on 7 October 2026 by Underdog, the on-device AI assistant made by Conway Research.[1][4][8] The model ships as a single 7.89 GB [GGUF](https://aiwiki.ai/wiki/gguf) file, against what Underdog gives as about 54 GB for the full-precision model, and loads in unmodified [llama.cpp](https://aiwiki.ai/wiki/llama_cpp).[1][2] Underdog says it targeted [tool calling](https://aiwiki.ai/wiki/function_calling) when shrinking the model, and its own tests put Saluki slightly ahead of the full-size Qwen3.8-27B on a 120-task tool-calling set while it gives up more ground on competition math.[1][2] The weights are published on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) under the Apache License 2.0, and Underdog calls the release a preview checkpoint.[1][2]

## Background

Underdog is a personal assistant for email, calendar, meeting notes, dictation and errands that runs its language models locally on the user's computer. Its website describes it as an invite-only beta and offers versions for Mac, iPhone and iPad, and Windows (labelled beta), with Linux and Android "on the way".[9] The product is made by Conway Research, and its founder is Sigil Wen.[8][10][16] On 2 October 2026 [Andreessen Horowitz](https://aiwiki.ai/wiki/andreessen_horowitz) published a post saying it was leading Conway Research's first round of funding; the post describes Underdog as Conway's first product.[8] Underdog's website says that first round was raised in April 2026, and it also lists [Khosla Ventures](https://aiwiki.ai/wiki/khosla_ventures), Hummingbird and other firms among its backers.[9]

On the Saluki launch page Underdog describes itself as "an application-first company" that cares about "building the best products using the best open source models".[1] Saluki is one of several models the company distributes through its app. Its website presents a 4B model called Woof as "the fast one", Saluki 27B as "the light one", and Underdog 27B, a separately post-trained derivative of Qwen 3.8 27B described in a 4 October 2026 post, as "the smart one".[9][11]

## Lineage and construction

Saluki is not a quantization made from scratch. Its NOTICE file states that it is a modified version of two Apache 2.0 works, the [Qwen](https://aiwiki.ai/wiki/qwen) team's Qwen3.8-27B and ISTA-DASLab's Qwen3.8-27B-GSQ-RCO-GGUF release, and that it "is built on" the file `Qwen3.8-27B-GSQ-RCO-IQ2_XS.gguf` from the latter.[3] The launch page likewise thanks ISTA-DASLab "for the 2-bit release Saluki is built on".[1]

ISTA-DASLab is the Deep Algorithms and Systems Lab at the Institute of Science and Technology Austria. Its Qwen3.8-27B repository, created on 28 August 2026, offers four non-uniform GGUF quantizations in which each tensor gets its own quantization type.[5] Two methods from the lab produce them:

| Method | What it does | Paper |
|---|---|---|
| GSQ (Gumbel-Softmax Quantization) | Post-training scalar quantization that learns grid assignments and per-group scales through a Gumbel-Softmax relaxation | Dadgarnia et al., arXiv 2604.18556 (April 2026)[6] |
| RCO (Riemannian Constrained Optimization) | Chooses a quantization type for each tensor under an exact total size budget, treating the budget as a Riemannian manifold | Helcig and Alistarh, arXiv 2605.00649 (May 2026)[7] |

The smallest ISTA-DASLab file, IQ2_XS, averages 2.50 bits per weight and measures 8.4 GB.[5] Saluki's file is about 0.5 GB smaller. Underdog's model card says the model was "tuned to keep tool calling intact", but the company had not published its training or quantization procedure at launch; the launch page says "We'll be publishing our full training soon."[1][2] BlockBeats summarized the release the same way: ISTA-DASLab had brought the model to about 8.4 GB and Underdog reduced it further while optimizing for tool calling.[12]

## Release and files

Underdog announced Saluki on X and on its website on 7 October 2026 (US time; the post is timestamped 02:27 UTC on 8 October).[4][1] The Hugging Face repository `ConwayResearch/Underdog-Saluki-27B-1.0` was created at 01:47 UTC on 8 October. A "Vision add-on" commit followed about 14 minutes later, and an Ollama section was added to the model card on 15 October.[13]

| Item | Detail |
|---|---|
| Developer | Underdog (Conway Research)[3][10] |
| Base model | Qwen3.8-27B; quantized from ISTA-DASLab's GSQ-RCO IQ2_XS GGUF[2][3] |
| Main file | `Underdog-Saluki-27B-1.0-IQ2-mix.gguf`, 7,898,369,152 bytes (7.89 GB)[2] |
| Full-precision size | About 54 GB[1][2] |
| Context length (GGUF metadata) | 262,144 tokens[2] |
| Vision add-on | Optional `mmproj` projector files, 928 MB (F16) or 629 MB (Q8_0)[2] |
| Runtime | Stock llama.cpp and apps built on it; model card also gives Ollama instructions[2] |
| License | Apache License 2.0[2][3] |
| Status | Preview checkpoint[1] |

The launch page says Saluki runs on a 16 GB laptop and describes the release as "text in, text out", adding that it "doesn't take images".[1] The model card, as updated after launch, lists the main file as text only and describes vision as an add-on: users who want image input download one of the projector files and pass it to llama-server with `--mmproj`.[2]

For general use the model card recommends leaving Qwen3.8's thinking mode on, with temperature 0.6, top_p 0.95 and top_k 20. For fast, direct tool calls it recommends turning thinking off and using temperature 0.[2] The card reports that chat, thinking and tool calling worked when tested with [Ollama](https://aiwiki.ai/wiki/ollama) 0.40.4 on Apple silicon.[2]

## Benchmarks

All benchmark figures below come from Underdog. The company ran Saluki in stock llama.cpp and ran each comparison model "in its own runtime", which it describes as that model's best case.[1] It cautions that "120 tasks is a modest test" and that part of any small lead is likely run-to-run variation.[1][2]

### Underdog Bench

Underdog Bench is Underdog's tool-calling test: 120 tasks from the [Berkeley Function Calling Leaderboard](https://aiwiki.ai/wiki/bfcl) (BFCL v4), 24 in each of five kinds of tool use, chosen and frozen on 27 September 2026 before any model was tested.[1][2] Runs used thinking off and temperature 0.[2]

| Model | File size | Tasks passed (of 120) |
|---|---|---|
| Underdog "Main" (a second Underdog build) | 7.9 GB | 91 |
| Underdog Saluki 27B | 7.89 GB | 88 |
| Mia 2-bit (EXL3, 2.0 bpw) | 9.1 GB | 85 |
| Qwen3.8-27B, full size | About 54 GB | 84 |
| ISTA 2-bit (IQ2_XS) | 8.4 GB | 76 |
| Ternary Bonsai 2 | 5.95 GB | 70 |

Source: Underdog launch page.[1]

Saluki also solved 76 of the 84 tasks the full-size model solved. Underdog highlights the two task types about choosing a function, where Saluki passed 47 of 48 against 31 for Mia 2-bit.[1] The company states that it does not claim its builds are smarter than the original model.[1]

### Parallel tool calls

On 100 parallel-calling tasks from BFCL, run with thinking off, Saluki passed 42 under strict scoring and 55 when small formatting slips were forgiven. The full-size model passed 35 under both scores, and Mia 2-bit passed 69 and 70.[1] Underdog says Mia "is the clear leader here". It notes that the full model, run with thinking off, often makes one call and stops when a task needs several. In 23 tasks Saluki's reply contained no call the checker could read.[1]

### Other evaluations

| Benchmark | Saluki | Full-size Qwen3.8-27B |
|---|---|---|
| [SWE-bench Verified](https://aiwiki.ai/wiki/swe_bench), issues fixed of 50 | 30 | 33 |
| [IFEval](https://aiwiki.ai/wiki/ifeval) (prompt-loose) | 93.5 | 91.5 (public score) |
| [IFBench](https://aiwiki.ai/wiki/ifbench) (prompt-loose) | 72.7 | 71.0 (public score) |
| MBPP+ | 78.0 | 83.9 (public score) |
| [MuSR](https://aiwiki.ai/wiki/musr) | 67.5 | 79.6 (public score) |
| [AIME 2025](https://aiwiki.ai/wiki/aime_2025) (avg@4) | 79.2 | 96.7 (public score) |
| AIME 2026 (avg@4) | 80.0 | 94.6 (public score) |

Source: Saluki model card.[2] "Public" scores come from a different harness than Underdog's.[2]

The launch page adds that 27 of Saluki's 30 SWE-bench fixes were issues the full model also fixed. On WebWalkerQA, a web-browsing test of 150 questions, Saluki answered 60 correctly; the full model was not run on that test.[1] On IFEval, Underdog notes that the public full-model figure does not say whether it is loose or strict, and that Saluki scored 90.9 on the strict measure.[1]

Underdog's headline claim is that Saluki keeps 96% of the full model's benchmark performance, averaged across nine benchmarks.[2][9] The company also says Saluki keeps 82 to 85% of the full model's score on competition math, which matches the two AIME rows above.[1][2]

## Limitations

Underdog lists these limitations on the launch page and model card:[1][2]

- On competition math it keeps only about 82 to 85% of the full model's AIME scores. The launch page describes the model as "built for agents, not math olympiads".
- Among instruction-following tasks, it is weakest on letter-level puzzles such as palindromes, vowel rules and alphabetical order.
- About a fifth of parallel-call replies contain small formatting slips, so Underdog recommends a lenient tool-call parser.
- With thinking on, the model often reasons at length before answering.
- The main file is text only; image input needs the separate vision file.

## Reception

Trade and crypto-industry news sites covered the release within days, reporting Underdog's figures as self-reported. BlockBeats called the tool-calling results "self-tested" and noted that performance "varied widely across different tasks".[12] AlphaSignal reported that Saluki completed more tool-calling tasks than the full-size model "according to Conway's tests".[14] On 15 October 2026 the Hugging Face API reported 344,177 downloads for the repository, a rolling 30-day count that covered the whole period since the repository was created.[15][13]

## See also

- [Qwen3.8](https://aiwiki.ai/wiki/qwen3_8)
- [Quantization](https://aiwiki.ai/wiki/quantization)
- [GGUF](https://aiwiki.ai/wiki/gguf)
- [llama.cpp](https://aiwiki.ai/wiki/llama_cpp)
- [Function calling](https://aiwiki.ai/wiki/function_calling)
- [On-device AI](https://aiwiki.ai/wiki/on_device_ai)

## References

1. Underdog. "Meet Underdog Saluki 27B." 7 October 2026. https://underdog.ai/saluki (accessed 15 October 2026).
2. ConwayResearch. "Underdog Saluki 27B 1.0" (model card). Hugging Face. https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0 (accessed 15 October 2026).
3. ConwayResearch. "NOTICE." Underdog-Saluki-27B-1.0 repository, Hugging Face. https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0/blob/main/NOTICE (accessed 15 October 2026).
4. Underdog AI (@UnderdogAI). "Today we're releasing Underdog Saluki 27B." X, 8 October 2026 (UTC). https://x.com/UnderdogAI/status/2108021482983133395
5. ISTA-DASLab. "Qwen3.8-27B GSQ-RCO GGUFs" (model card). Hugging Face. https://huggingface.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF (accessed 15 October 2026).
6. Dadgarnia, A.; Tabesh, S.; Nikdan, M.; Helcig, M.; Kurtic, E.; Kleinegger, M.; Alistarh, D. "GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling." arXiv:2604.18556, 20 April 2026. https://arxiv.org/abs/2604.18556
7. Helcig, M.; Alistarh, D. "Model Compression with Exact Budget Constraints via Riemannian Manifolds." arXiv:2605.00649, 1 May 2026. https://arxiv.org/abs/2605.00649
8. Dixon, C.; Vasquez, G.; Harkavy, E. "Investing in Conway, the Creator of Underdog." Andreessen Horowitz, 2 October 2026. https://a16z.com/announcement/investing-in-conway-the-creator-of-underdog/
9. Underdog. "Underdog. The Most Loyal AI." (homepage). https://underdog.ai/ (accessed 15 October 2026).
10. Underdog. "Security." Updated 10 October 2026. https://underdog.ai/security (accessed 15 October 2026).
11. Underdog. "Meet Underdog 27B." 4 October 2026. https://underdog.ai/underdog-27b
12. BlockBeats. "Underdog compresses Qwen3.8-27B to 7.89GB: self-tested tool calling surpasses the original version." 8 October 2026. https://en.theblockbeats.news/flash/370754
13. ConwayResearch. "Underdog-Saluki-27B-1.0: commit history." Hugging Face. https://huggingface.co/ConwayResearch/Underdog-Saluki-27B-1.0/commits/main (accessed 15 October 2026).
14. AlphaSignal. "Conway Research Squeezes Saluki's 27B Model Into 8 GB and Beats the Original." 10 October 2026. https://alphasignal.ai/news/conway-research-squeezes-saluki-s-27b-model-into-8-gb-and-beats-the-original
15. Hugging Face. Model API record for ConwayResearch/Underdog-Saluki-27B-1.0. https://huggingface.co/api/models/ConwayResearch/Underdog-Saluki-27B-1.0 (accessed 15 October 2026).
16. Wen, S. "Own your AI, don't let it own you" (manifesto). Underdog, 1 October 2026. https://underdog.ai/manifesto