Swift-Qwen3.8-27b on AMD GPUs

57 results · 2 sources · 78 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	Swift-Qwen3.8-27bQ4_K_SRepository unresolved	llm-bench.io QualityFour sequential scenarios: agent workflow, code generation, role play, research	81.52 / 100LLM judge · Mean	246.88 tok/sPrefill · mean · per request	29.78 tok/sGeneration · mean · per request	GPU count unknownUnknown · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS Windows Engine LM Studio Backend Not reported Harness llm-benchmark 0.4.85+101 Prompt / output ? / ? tokens Context 80000 Concurrency 1 Date 2026-10-01 Full record →
View ReportCustom evaluation	Swift-Qwen3.8-27bMXFP6 E2M3 W6A8Model ref. ↗	GSM8KFirst 500 test questions; two few-shot seeds256K configured context	96.7 %Accuracy · n = 100 · Flexible exact match · Mean	—	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness lm-eval Prompt / output ? / ? tokens Context 262144 Concurrency Not reported Date 2026-09-16 Full record →
View Reportllm-bench.io	Swift-Qwen3.8-27bQ4_K_MWeights ↗	llm-bench.io QualityFour sequential scenarios: agent workflow, code generation, role play, research	87.44 / 100LLM judge · Mean	255.66 tok/sPrefill · mean · per request	32.89 tok/sGeneration · mean · per request	GPU count unknownUnknown · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS Windows 11 Home 25H2 Engine Unsloth Studio Backend Not reported Harness llm-benchmark 0.4.80+98 Prompt / output ? / ? tokens Context 80000 Concurrency 1 Date 2026-09-24 Full record →
View ReportCustom evaluation	Swift-Qwen3.8-27bMXFP6 E2M3 W6A8Model ref. ↗	GSM8KFirst 500 test questions; two few-shot seeds256K configured context	95.9 %Accuracy · n = 100 · Strict exact match · Mean	—	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness lm-eval Prompt / output ? / ? tokens Context 262144 Concurrency Not reported Date 2026-09-16 Full record →
View Reportllm-bench.io	Swift-Qwen3.8-27bQ4_K_MWeights ↗	llm-bench.io QualityFour sequential scenarios: agent workflow, code generation, role play, research	87.31 / 100LLM judge · Mean	606.05 tok/sPrefill · mean · per request	31.76 tok/sGeneration · mean · per request	GPU count unknownUnknown · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS Windows 11 Home 25H2 Engine Unsloth Studio Backend Not reported Harness llm-benchmark 0.4.80+98 Prompt / output ? / ? tokens Context 65536 Concurrency 1 Date 2026-09-24 Full record →
View ReportGeneric LLM	Swift-Qwen3.8-27bMXFP6 E2M3 W6A8Model ref. ↗	WikiTextwikitext-2-raw-v1 full test; 145 nonoverlapping 2048-token chunks1 concurrent request	7.0826 perplexityWord perplexity · Token-weighted perplexity · Lower is better	—	—	1 × R9700Single GPU · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness Custom prompt_logprobs perplexity collector Prompt / output ? / ? tokens Context Not reported Concurrency 1 Date 2026-09-16 Full record →
View Reportllm-bench.io	Swift-Qwen3.8-27bQ4_K_MWeights ↗	llm-bench.io QualityFour sequential scenarios: agent workflow, code generation, role play, research	86.81 / 100LLM judge · Mean	256.53 tok/sPrefill · mean · per request	33.2 tok/sGeneration · mean · per request	GPU count unknownUnknown · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS Windows 11 Home 25H2 Engine Unsloth Studio Backend Not reported Harness llm-benchmark 0.4.80+98 Prompt / output ? / ? tokens Context 80000 Concurrency 1 Date 2026-09-24 Full record →
View ReportCustom evaluation	Swift-Qwen3.8-27bMXFP6 E2M3 W6A8Model ref. ↗	GSM8KFirst 500 test questions; two few-shot seeds256K configured context	96.7 %Accuracy · n = 100 · Flexible exact match · Mean	—	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness lm-eval Prompt / output ? / ? tokens Context 262144 Concurrency Not reported Date 2026-09-16 Full record →
View Reportllm-bench.io	Swift-Qwen3.8-27bQ4_K_MWeights ↗	llm-bench.io QualityFour sequential scenarios: agent workflow, code generation, role play, research	84.66 / 100LLM judge · Mean	272.09 tok/sPrefill · mean · per request	34.83 tok/sGeneration · mean · per request	GPU count unknownUnknown · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS Windows 11 Home 25H2 Engine Unsloth Studio Backend Not reported Harness llm-benchmark 0.4.80+98 Prompt / output ? / ? tokens Context 80000 Concurrency 1 Date 2026-09-24 Full record →
View ReportCustom evaluation	Swift-Qwen3.8-27bMXFP6 E2M3 W6A8Model ref. ↗	GSM8KFirst 500 test questions; two few-shot seeds256K configured context	95.5 %Accuracy · n = 100 · Strict exact match · Mean	—	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness lm-eval Prompt / output ? / ? tokens Context 262144 Concurrency Not reported Date 2026-09-16 Full record →
View Reportllm-bench.io	Swift-Qwen3.8-27bQ4_K_MWeights ↗	llm-bench.io QualityFour sequential scenarios: agent workflow, code generation, role play, research	83.98 / 100LLM judge · Mean	622.41 tok/sPrefill · mean · per request	33.16 tok/sGeneration · mean · per request	GPU count unknownUnknown · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS Windows 11 Home 25H2 Engine Unsloth Studio Backend Not reported Harness llm-benchmark 0.4.80+98 Prompt / output ? / ? tokens Context 65536 Concurrency 1 Date 2026-09-24 Full record →
View ReportGeneric LLM	Swift-Qwen3.8-27bMXFP6 E2M3 W6A8Model ref. ↗	WikiTextwikitext-2-raw-v1 full test; 145 nonoverlapping 2048-token chunks1 concurrent request	7.0892 perplexityWord perplexity · Token-weighted perplexity · Lower is better	—	—	1 × R9700Single GPU · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness Custom prompt_logprobs perplexity collector Prompt / output ? / ? tokens Context Not reported Concurrency 1 Date 2026-09-16 Full record →
View Reportllm-bench.io	Swift-Qwen3.8-27bQ4_K_MWeights ↗	llm-bench.io QualityFour sequential scenarios: agent workflow, code generation, role play, research	80.26 / 100LLM judge · Mean	264.39 tok/sPrefill · mean · per request	34.55 tok/sGeneration · mean · per request	GPU count unknownUnknown · 24 GB VRAM/GPU Configuration Memory 24 GB VRAM/GPU OS Windows 11 Home 25H2 Engine Unsloth Studio Backend Not reported Harness llm-benchmark 0.4.80+98 Prompt / output ? / ? tokens Context 80000 Concurrency 1 Date 2026-09-24 Full record →
View ReportCustom evaluation	Swift-Qwen3.8-27bINT5 ParoQuant W5A8Model ref. ↗	GSM8KFirst 500 test questions; two few-shot seeds256K configured context	96.2 %Accuracy · n = 100 · Flexible exact match · Mean	—	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness lm-eval Prompt / output ? / ? tokens Context 262144 Concurrency Not reported Date Not reported Full record →
View ReportCustom evaluation	Swift-Qwen3.8-27bINT5 ParoQuant W5A8Model ref. ↗	GSM8KFirst 500 test questions; two few-shot seeds256K configured context	93.3 %Accuracy · n = 100 · Strict exact match · Mean	—	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness lm-eval Prompt / output ? / ? tokens Context 262144 Concurrency Not reported Date Not reported Full record →
View ReportCustom evaluation	Swift-Qwen3.8-27bINT5 ParoQuant W5A8Model ref. ↗	GSM8KFirst 500 test questions; two few-shot seeds256K configured context	97 %Accuracy · n = 100 · Flexible exact match · Mean	—	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness lm-eval Prompt / output ? / ? tokens Context 262144 Concurrency Not reported Date Not reported Full record →
View ReportCustom evaluation	Swift-Qwen3.8-27bINT5 ParoQuant W5A8Model ref. ↗	GSM8KFirst 500 test questions; two few-shot seeds256K configured context	93.5 %Accuracy · n = 100 · Strict exact match · Mean	—	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness lm-eval Prompt / output ? / ? tokens Context 262144 Concurrency Not reported Date Not reported Full record →
View ReportBetterBench	Swift-Qwen3.8-27bINT5 ParoQuant W5A8Model ref. ↗	Token speedWeighted eight-category mix	—	—	160.7 tok/sGeneration · per request	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness BetterBench Prompt / output ? / ? tokens Context 262144 Concurrency 1 Date 2026-09-15 Full record →
View ReportBetterBench	Swift-Qwen3.8-27bINT5 ParoQuant W5A8Model ref. ↗	Token speed	136.9 tok/sThroughput · mean · aggregate	—	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness BetterBench Prompt / output ? / ? tokens Context 262144 Concurrency 1 Date 2026-09-15 Full record →
View ReportBetterBench	Swift-Qwen3.8-27bINT5 ParoQuant W5A8Model ref. ↗	Token speed	252.6 tok/sThroughput · mean · aggregate	—	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness BetterBench Prompt / output ? / ? tokens Context 262144 Concurrency 2 Date 2026-09-15 Full record →
View ReportBetterBench	Swift-Qwen3.8-27bINT5 ParoQuant W5A8Model ref. ↗	Token speed	370.9 tok/sThroughput · mean · aggregate	—	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness BetterBench Prompt / output ? / ? tokens Context 262144 Concurrency 4 Date 2026-09-15 Full record →
View ReportBetterBench	Swift-Qwen3.8-27bINT5 ParoQuant W5A8Model ref. ↗	Token speed	427.3 tok/sThroughput · mean · aggregate	—	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness BetterBench Prompt / output ? / ? tokens Context 262144 Concurrency 8 Date 2026-09-15 Full record →
View ReportBetterBench	Swift-Qwen3.8-27bINT5 ParoQuant W5A8Model ref. ↗	Token speed	—	3,027 tok/sPrefill · per request	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness BetterBench Prompt / output ? / ? tokens Context 262144 Concurrency 1 Date 2026-09-15 Full record →
View ReportBetterBench	Swift-Qwen3.8-27bINT5 ParoQuant W5A8Model ref. ↗	Token speed	—	3,222 tok/sPrefill · per request	—	2 × R9700Tensor parallel · 32 GB VRAM/GPU Configuration Memory 32 GB VRAM/GPU OS Not reported Engine vLLM Radiance Backend ROCm / HIP Harness BetterBench Prompt / output ? / ? tokens Context 262144 Concurrency 1 Date 2026-09-15 Full record →

Quantizations and software

Quantizations: INT5 ParoQuant W5A8, MXFP6 E2M3 W6A8, Q4_K_M, Q4_K_S
Engines: LM Studio, Unsloth Studio, vLLM Radiance

Explore all 57 results →

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

GPU benchmark pages: R9700 · RX 7900 XTX