Qwen3.5-9B on AMD GPUs

241 results · 15 sources · 410 measurements

Grouped by reported model identity. Exact weights, quantizations and software can differ. How results are reviewed

Documented results

24 selected tests


View Reportllm-bench.io	Qwen3.5-9BQ4_K_MModel ref. ↗	llm-bench.io QualityFour sequential scenarios: agent workflow, code generation, role play, research	70.68 / 100LLM judge · Mean	—	64.64 tok/sGeneration · mean · per request	GPU count unknownUnknown · 16 GB VRAM/GPU Configuration Memory 16 GB VRAM/GPU OS openSUSE Tumbleweed Engine Ollama Backend Not reported Harness llm-benchmark 0.4.36+97 Prompt / output ? / ? tokens Context 65536 Concurrency 1 Date 2026-08-15 Full record →
View Reportllm-bench.io	Qwen3.5-9BQ4_K_MModel ref. ↗	llm-bench.io QualityFour sequential scenarios: agent workflow, code generation, role play, research	68.5 / 100LLM judge · Mean	—	18.36 tok/sGeneration · mean · per request	GPU count unknownUnknown · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS Windows Engine Ollama Backend Not reported Harness llm-benchmark 0.3.0+52 Prompt / output ? / ? tokens Context 65536 Concurrency 1 Date 2026-04-13 Full record →
View Reportllm-bench.io	Qwen3.5-9BQ4_K_MModel ref. ↗	llm-bench.io QualityFour sequential scenarios: agent workflow, code generation, role play, research	62.5 / 100LLM judge · Mean	1,029.22 tok/sPrefill · mean · per request	70.18 tok/sGeneration · mean · per request	GPU count unknownUnknown · 16 GB VRAM/GPU Configuration Memory 16 GB VRAM/GPU OS Windows 11 Pro 26H2 Engine LM Studio Backend Not reported Harness llm-benchmark 0.4.96+102 Prompt / output ? / ? tokens Context 65536 Concurrency 1 Date 2026-10-17 Full record →
View ReportGeneric LLM	Qwen3.5-9BUD-Q4_K_XLModel ref. ↗	HumanEvalFull 164 problems	67.07 %pass@1 · n = 164 · Executable tests	—	37.98 tok/sGeneration · mean · per request	1 × RX 9070Single GPU · 16 GB VRAM/GPU Configuration Memory 16 GB VRAM/GPU OS Windows / WSL2 Engine llama.cpp Backend ROCm / HIP Harness custom OpenAI API HumanEval runner Prompt / output ? / ? tokens Context 16384 Concurrency Not reported Date Not reported Full record →
View ReportCustom evaluation	Qwen3.5-9BQuant not reportedModel ref. ↗	Local LLM custom suite4b · Math / math_basic	100 %Score · 1 / 1 · Custom answer parser 60.92 tok/sThroughput · per request	—	—	1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU Configuration Memory 16 GB VRAM/GPU OS Ubuntu 24.04 LTS Engine Ollama Backend ROCm / HIP Harness benchmark_v4b.py Prompt / output ? / ? tokens Context Not reported Concurrency 1 Date 2026-03-19 Full record →
View ReportCustom evaluation	Qwen3.5-9BQ8_0Model ref. ↗	Custom C++ corpus perplexity600KB C++ corpus2K configured context	1.5258 perplexityScore · Perplexity · Lower is better	—	—	1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU Configuration Memory 16 GB VRAM/GPU OS NixOS Engine llama.cpp TurboQuant fork Backend ROCm / HIP Harness llama-perplexity Prompt / output ? / ? tokens Context 2048 Concurrency Not reported Date Not reported Full record →
View ReportLocalMaxxing	Qwen3.5-9BMQ4Model ref. ↗	Token speed	—	754.54 tok/sPrefill · unknown	576.85 tok/sGeneration · unknown	1 × RX 7900 XTXSingle GPU · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS Not reported Engine hipfire Backend Not reported Harness LocalMaxxing community speed test; exact client where reported Prompt / output 27 / 256 tokens Context 4096 Concurrency Not reported Date 2026-05-20 Full record →
View ReportGeneric LLM	Qwen3.5-9BUD-Q4_K_XLModel ref. ↗	Token speed	—	1,851 tok/sPrefill · per request	83.82 tok/sGeneration · per request	1 × RX 7900 XTSingle GPU · 20 GB VRAM/GPU Configuration Memory 20 GB VRAM/GPU OS Debian GNU/Linux13.3 / kernel6.18.2-zen4 Engine llama.cpp Backend Vulkan Harness Author repeated llama-server prompt loop Prompt / output ? / ? tokens Context 262144 Concurrency 1 Date Not reported Full record →
View ReportLocalMaxxing	Qwen3.5-9BMQ4Model ref. ↗	Token speed	—	1,160.07 tok/sPrefill · mean · unknown	371.8 tok/sGeneration · mean · unknown	1 × R9700Single GPU · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Ubuntu 24.04 Engine hipfire Backend rocm Harness LocalMaxxing community speed test; exact client where reported Prompt / output 27 / 256 tokens Context 4096 Concurrency Not reported Date 2026-05-20 Full record →
View ReportLocalMaxxing	Qwen3.5-9BIQ4_NLModel ref. ↗	Token speed	—	700.1 tok/sPrefill · median · unknown	38.6 tok/sGeneration · median · unknown	1 × Ryzen AI Max+ 395Single GPU · Unified memory Configuration Memory Unified memory OS Ubuntu 24.04 Engine llama.cpp Backend vulkan Harness LocalMaxxing community speed test; exact client where reported Prompt / output 1185 / 512 tokens Context 8192 Concurrency Not reported Date 2026-08-12 Full record →
View Reportllama-bench	Qwen3.5-9BUD-Q4_K_XLModel ref. ↗Conflicting evidence	Token speed	—	2,115.89 tok/sPrefill · mean · per request	—	1 × RX 9070Single GPU · 16 GB VRAM/GPU Configuration Memory 16 GB VRAM/GPU OS Windows 11 Engine llama.cpp Backend Vulkan Harness llama-bench Prompt / output 8192 / 0 tokens Context Not reported Concurrency Not reported Date 2026-09-01 Full record →
View Reportllama-bench	Qwen3.5-9BQ4_K_MModel ref. ↗	Token speed	—	1,845.34 tok/sPrefill · mean · per request	—	1 × RX 9060 XTSingle GPU · 16 GB VRAM/GPU Configuration Memory 16 GB VRAM/GPU OS CachyOS / kernel 7.1.3-2-cachyos Engine llama.cpp Backend ROCm / HIP Harness llama-bench Prompt / output 512 / 0 tokens Context Not reported Concurrency Not reported Date Not reported Full record →
View ReportGeneric LLM	Qwen3.5-9BQ8_0Model ref. ↗	Token speed	—	—	58.37 tok/sGeneration · mean · per request	1 × R9700Single GPU · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Ubuntu 24.04 Engine llama.cpp Backend ROCm 7.0.0 Harness Not reported Prompt / output ? / 128 tokens Context Not reported Concurrency Not reported Date 2026-05-14 Full record →
View Reportllama-bench	Qwen3.5-9BBF16Model ref. ↗	Token speed2,048 input tokens · 16,384 existing KV tokens	—	2,455.97 tok/sPrefill · mean · per request	—	1 × R9700Single GPU · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Fedora 42 — Linux 6.15.9-201.fc42.x86_64 (Sat Aug 2 11:37:34 UTC 2025) Engine llama.cpp Backend rocm6_4_4 Harness llama-bench Prompt / output 2048 / ? tokens Context Not reported Concurrency Not reported Date Not reported Full record →
View ReportGeneric LLM	Qwen3.5-9BQ4_K_MModel ref. ↗	Token speed	—	—	83.4 tok/sGeneration · mean · per request	1 × R9700Single GPU · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Ubuntu 24.04.4 LTS Engine llama.cpp Backend ROCm 7.2.4 Harness Not reported Prompt / output ? / ? tokens Context 32768 Concurrency Not reported Date 2026-09-23 Full record →
View Reportllm-bench.io	Qwen3.5-9BQ4_K_MModel ref. ↗	llm-bench.io QualityFour sequential scenarios: agent workflow, code generation, role play, research	66.96 / 100LLM judge · Mean	1,529.02 tok/sPrefill · mean · per request	71.66 tok/sGeneration · mean · per request	GPU count unknownUnknown · 16 GB VRAM/GPU Configuration Memory 16 GB VRAM/GPU OS Windows 11 Pro 25H2 Engine Ollama Backend Not reported Harness llm-benchmark 0.4.59+97 Prompt / output ? / ? tokens Context 8192 Concurrency 1 Date 2026-09-18 Full record →
View Reportllm-bench.io	Qwen3.5-9BQ4_K_MModel ref. ↗	llm-bench.io QualityFour sequential scenarios: agent workflow, code generation, role play, research	67.88 / 100LLM judge · Mean	—	16.57 tok/sGeneration · mean · per request	GPU count unknownUnknown · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS Windows Engine Ollama Backend Not reported Harness llm-benchmark 0.4.14+90 Prompt / output ? / ? tokens Context 16384 Concurrency 1 Date 2026-04-17 Full record →
View ReportGeneric LLM	Qwen3.5-9BUD-Q4_K_XLModel ref. ↗	HumanEvalFull 164 problems	67.07 %pass@1 · n = 164 · Executable tests	—	39.8 tok/sGeneration · mean · per request	1 × RX 9070Single GPU · 16 GB VRAM/GPU Configuration Memory 16 GB VRAM/GPU OS Windows / WSL2 Engine llama.cpp Backend ROCm / HIP Harness custom OpenAI API HumanEval runner Prompt / output ? / ? tokens Context 16384 Concurrency Not reported Date Not reported Full record →
View ReportCustom evaluation	Qwen3.5-9BQuant not reportedModel ref. ↗	Local LLM custom suite4b · Math / math_discount	100 %Score · 1 / 1 · Custom answer parser 60.29 tok/sThroughput · per request	—	—	1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU Configuration Memory 16 GB VRAM/GPU OS Ubuntu 24.04 LTS Engine Ollama Backend ROCm / HIP Harness benchmark_v4b.py Prompt / output ? / ? tokens Context Not reported Concurrency 1 Date 2026-03-19 Full record →
View ReportCustom evaluation	Qwen3.5-9BQ8_0Model ref. ↗	Custom C++ corpus perplexity600KB C++ corpus2K configured context	1.5281 perplexityScore · Perplexity · Lower is better	—	—	1 × RX 9070 XTSingle GPU · 16 GB VRAM/GPU Configuration Memory 16 GB VRAM/GPU OS NixOS Engine llama.cpp TurboQuant fork Backend ROCm / HIP Harness llama-perplexity Prompt / output ? / ? tokens Context 2048 Concurrency Not reported Date Not reported Full record →
View ReportLocalMaxxing	Qwen3.5-9BMQ4Model ref. ↗	Token speed	—	—	123.11 tok/sGeneration · unknown	1 × RX 7900 XTXSingle GPU · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS Not reported Engine hipfire Backend Not reported Harness LocalMaxxing community speed test; exact client where reported Prompt / output 27 / 156 tokens Context 4096 Concurrency Not reported Date 2026-05-16 Full record →
View ReportGeneric LLM	Qwen3.5-9BUD-Q4_K_XLModel ref. ↗	Token speed	—	1,129 tok/sPrefill · mean · per request	82.41 tok/sGeneration · mean · per request	1 × RX 7900 XTSingle GPU · 20 GB VRAM/GPU Configuration Memory 20 GB VRAM/GPU OS Debian GNU/Linux13.3 / kernel6.18.2-zen4 Engine llama.cpp Backend Vulkan Harness Author repeated llama-server prompt loop Prompt / output ? / ? tokens Context 262144 Concurrency 1 Date Not reported Full record →
View ReportLocalMaxxing	Qwen3.5-9BQ4_K_MModel ref. ↗	Token speed	—	3,288.79 tok/sPrefill · unknown	90.49 tok/sGeneration · unknown	1 × R9700Single GPU · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Fedora Linux 44 (kernel 7.2.7-200.fc44.x86_64) Engine llama.cpp Backend vulkan Harness LocalMaxxing community speed test; exact client where reported Prompt / output 512 / 128 tokens Context 256 Concurrency Not reported Date 2026-10-01 Full record →
View ReportLocalMaxxing	Qwen3.5-9BIQ4_NLModel ref. ↗	Token speed	—	684.5 tok/sPrefill · median · unknown	37.5 tok/sGeneration · median · unknown	1 × Ryzen AI Max+ 395Single GPU · Unified memory Configuration Memory Unified memory OS Ubuntu 24.04 Engine llama.cpp Backend vulkan Harness LocalMaxxing community speed test; exact client where reported Prompt / output 1185 / 512 tokens Context 8192 Concurrency Not reported Date 2026-08-12 Full record →

Quantizations and software

Quantizations: BF16, IQ4_NL, MQ4, Q4_K_M, Q8_0, UD-Q4_K_XL, UD-Q5_K_XL, UD-Q6_K_XL, UD-Q8_K_XL
Engines: hipfire, llama.cpp, llama.cpp TurboQuant fork, LM Studio, Ollama

Explore all 241 results →

This is a selection across sources, not a ranking or a matched experiment. Reading benchmark reports

GPU benchmark pages: R9700 · RX 9070 XT · RX 9070 · RX 9060 XT · RX 7900 XTX · RX 7900 XT · RX 7900 GRE · RX 7800 XT · Ryzen AI Max+ 395