Underdog Saluki 27B 1.0
Qwen3.8-27B compressed into 7.89 GB with tool calling kept intact — a 2-bit GGUF that runs serious agents on ordinary hardware.
About
Underdog Saluki 27B 1.0 is a 2-bit quantized build of Qwen3.8-27B (IQ2-mix, 7.89 GB) tuned to preserve agentic behavior. On the vendor's Underdog Bench it keeps 96% of the full model's capability across 9 benchmarks, scoring 88 vs 84 on tool calling and 42 vs 35 on parallel tool calls — numbers that come from Conway Research's own harness, not independent testing. It runs on stock llama.cpp with an OpenAI-compatible server, keeps Qwen3.8's switchable thinking mode, and offers an optional mmproj vision add-on for image input. Conway positions it for local agents, coding assistants, and instruction-heavy workloads where a full 54 GB model is impractical. Released 2026-10-08 under Apache 2.0, it gathered about 371k downloads and 513 likes in its first week on Hugging Face.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Underdog Bench tool calling: 88 vs full model 84 — provider-reported
- Parallel tool calls: 42 vs 35 — provider-reported
- Average retention 96% across 9 benchmarks vs full Qwen3.8-27B — provider-reported
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 5/10
Scored 7/10 (METRICS tier, metrics-v1). Evidence: ~371k downloads in 9 days, strong signal of real demand; Apache 2.0; standard GGUF for stock llama.cpp with documented quickstart; quant methodology (GSQ-RCO via ISTA-DASLab) disclosed. Docked for: (1) all retention claims come from the vendor's own bench, no independent eval seen; (2) new small company, single release so far. License: Apache-2.0.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 8.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · HF: 513 likes, ~371k downloads in 9 days, created 2026-10-08
- · manual review: quant base and method disclosed; standard GGUF for stock llama.cpp
- · docked: all retention claims from the vendor's own bench; new small company, single release so far
More in the qwen3 family
- Qwen3-32B-AWQ
- Qwen3-Next-80B-A3B
- Qwen3-235B-A22B
- Qwen3-235B-A22B-2507
- Qwen3-4B-Instruct-2507
- Qwen3-8B
- Qwen3.5
- Qwen3-235B-A22B-Instruct-2507
- Qwen3 235B
- Qwen3-235B-A22B-Thinking-2507
- Qwen3-Omni-Flash
- Qwen3.6-27B
- Qwen3-32B
- Qwen3-Max
- Qwen3.8-Flash-Next
- Qwen3-Coder-Plus
- Qwen3
- Qwen3-0.6B
- Qwen3-4B
- Qwen3-Coder-Next
- Qwen3-4B-2507
- Qwen3-14B
- Qwen3-VL-30B-A3B
- Qwen3-1.7B
- Qwen3-Omni-30B-A3B