Underdog Saluki 27B 1.0 has 26.896 billion total parameters. The selected IQ2 mixed GGUF file is 7.898 GB.
Data snapshot: . 7 published tests in this catalog.
Model file and details
Provider
Conway Research
Total parameters
26,895,998,464
Exact file size
7,898,369,152 bytes
Quantization / precision
IQ2 mixed
Weight repository
ConwayResearch/Underdog-Saluki-27B-1.0
Original model repository
ConwayResearch/Underdog-Saluki-27B-1.0
Weight revision
5acfd9218fa54731f237eea8b3243d430dd85a69
Model revision
5acfd9218fa54731f237eea8b3243d430dd85a69
Runtime notes
Confirm model architecture and quantization support in your runner.
Recorded release date
2026-10-08
Date evidence
Conway Research released the Apache-2.0 Underdog Saluki 27B 1.0 repository and exact IQ2-mixed GGUF on 8 October 2026. The selected 7,898,369,152-byte file is a text-only, 26,895,998,464-parameter Qwen3.8-27B derivative intended for stock llama.cpp; the publisher added Ollama instructions at revision 5acfd9218fa54731f237eea8b3243d430dd85a69 on 15 October. Vision requires a separate 629 MB Q8_0 or 928 MB F16 projector, and the weight size excludes runtime, KV cache, context and output headroom.
These scores were compiled from the linked sources. These published scores were not independently tested here. Different prompts, precision, thinking budgets and tools can change results.
IFEval · reported
93.5 / 100
Following precise instructions. These cards report IFEval without consistently specifying prompt-level versus instruction-level scoring. Treat this view as approximate.
Benchmark efficiency: 11.838 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Conway Research-reported IFEval prompt-loose score for the exact selected IQ2-mixed GGUF in its stock llama.cpp evaluation harness.
Publisher-run result, not independent validation. The card does not fully enumerate prompts, shots, sampling, repeats or llama.cpp revision for this row. Its public full-size comparison comes from a different harness and should not be treated as a controlled head-to-head result.
Competition math from AIME 2025. Published answer accuracy.
Benchmark efficiency: 10.027 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Conway Research-reported AIME 2025 average-at-4 score for the exact selected IQ2-mixed GGUF in its stock llama.cpp evaluation harness.
Publisher-run result, not independent validation. The card does not fully specify prompt, sampling, answer extraction, reasoning budget or run seeds. The public full-size comparison uses a different harness.
Competition math from AIME 2026. Published answer accuracy.
Benchmark efficiency: 10.129 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Conway Research-reported AIME 2026 average-at-4 score for the exact selected IQ2-mixed GGUF in its stock llama.cpp evaluation harness.
Publisher-run result, not independent validation. The card does not fully specify prompt, sampling, answer extraction, reasoning budget or run seeds. The public full-size comparison uses a different harness.
Basic Python programming with extra tests. EvalPlus pass@1 on the sanitized 378-problem set.
Benchmark efficiency: 9.875 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Conway Research-reported MBPP+ score for the exact selected IQ2-mixed GGUF in its stock llama.cpp evaluation harness.
Publisher-run result, not independent validation. Prompting, pass metric, sampling and harness details are incomplete in the card; the public full-size comparison uses a different harness.
Instruction following: prompt-level loose accuracy. Separate from IFEval and other IFBench metrics.
Benchmark efficiency: 9.204 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Conway Research-reported IFBench prompt-loose score for the exact selected IQ2-mixed GGUF in its stock llama.cpp evaluation harness.
Publisher-run result, not independent validation. The card does not fully enumerate prompts, sampling, repeats or llama.cpp revision for this row. Its public full-size comparison uses a different harness.
Publisher-reported passed-task count on a frozen 120-task subset derived from BFCL v4. This is a limited tool-calling panel, not the full official leaderboard.
Benchmark efficiency: 11.142 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Conway Research reports 88 passed tasks out of a 120-task BFCL v4-derived panel frozen before testing. The exact selected IQ2-mixed GGUF ran in stock llama.cpp with thinking disabled and temperature 0. The publisher says its full-size Qwen3.8-27B comparison used the same harness and settings.
Publisher-run subset, not an independent result or the full official BFCL v4 leaderboard. Task selection, harness, checker, prompt formatting and run variance affect the count; the publisher calls 120 tasks a modest test. The top-card shorthand '88' is retained here as 88 of 120, not 88 percent. Apache-2.0 license.
Publisher-reported passed-task count on 100 BFCL v4 parallel-call tasks using the official checker. This is a limited publisher-run panel.
Benchmark efficiency: 5.318 published score points per GB of selected weights. This ratio is not an inference-speed measure; check the test settings to see whether the score covers the selected quantization.
Test settings: Conway Research reports 42 passed tasks on 100 BFCL v4 parallel-call tasks using the official checker, with thinking disabled, on the exact selected IQ2-mixed GGUF in stock llama.cpp.
Publisher-run limited panel, not an independent validation or full leaderboard score. The card warns that about one fifth of parallel-call replies have small formatting slips. Harness, prompt template, checker and run variance affect the count.
Context and version evidence: The repository and file name identify the selected model. Revision hashes are shown above when recorded. Tested context length, full hardware configuration and run date are not consistently recorded in this snapshot. Only settings explicitly stated in each benchmark record are known.
The 7.898 GB figure covers model weights in this file. It excludes the context cache, runtime, operating system and any separate multimodal files. It is not a measured RAM requirement or a token-generation speed.
These complete configurations pass the initial weight-size check for the selected 7.90 GB file. Allow extra memory for context and software. Check the model’s runtime notes and your inference software before buying.
Listing evidence: 2026-09-25. Amazon.com base price visible in an owner-supplied screenshot reviewed 25 Sep 2026; capture time not supplied. Dated reference, not a live quote. Excludes card signup offers, coupons, tax and shipping.
Listing evidence: 2026-10-03. Apple US launch price announced 22 September 2026: USD 1,699. Dated manufacturer reference, not an Amazon offer; tax and shipping may differ.
Listing evidence: 2026-09-24. GMKtec, EVO-X2 / 64GB / 1TB, recorded 22 Sep 2026; exact ASIN capacity identified 24 Sep. This is a dated reference, not a current Amazon price; tax, shipping and bundle contents may differ.
Memory matches are estimates, not tested compatibility or speed. As an Amazon Associate I earn from qualifying purchases.
Confirm model architecture and quantization support in your runner. The selected weights use 7.898 GB; this excludes context cache, runtime buffers and other programs.
These are the first listed tiers that pass our reserve policy, not measured minimum requirements. System RAM and VRAM remain separate. Use the memory tool to change the reserve for your context and runner.