The performance of the original model Qwen3.8-27B has also been measured across various benchmarks. A partial excerpt of the text performance comparison table is as follows:<\/p>\n
\n
\u2192 Scroll horizontally to see all columns<\/p>\n
\n\n\n| <\/th>\n | Qwen3.8-27B<\/th>\n | Qwen3.6-27B<\/th>\n | Qwen3.7-Plus<\/th>\n | Muse Glimmer-30B<\/th>\n | Opus4.6 Max<\/th>\n<\/tr>\n<\/thead>\n |
\n\n| Coding<\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n<\/tr>\n |
\n| Agentic terminal coding Terminal Bench 2.1 (Terminus)<\/td>\n | 73.0<\/td>\n | 63.4<\/td>\n | 64.0<\/td>\n | 51.7<\/td>\n | 78.2<\/td>\n<\/tr>\n |
\n| Agentic coding SWE-bench Pro<\/td>\n | 61.7<\/td>\n | 53.5<\/td>\n | 57.6<\/td>\n | 51.2<\/td>\n | 53.4<\/td>\n<\/tr>\n |
\n| Repo-level code generation NL2Repo-Bench<\/td>\n | 42.3<\/td>\n | 36.2<\/td>\n | 41.1<\/td>\n | —<\/td>\n | 47.6<\/td>\n<\/tr>\n |
\n| Agentic coding DeepSWE 1.1<\/td>\n | 42.2<\/td>\n | 13.3<\/td>\n | 14.2<\/td>\n | —<\/td>\n | —<\/td>\n<\/tr>\n |
\n| Software engineering QwenSWEBench<\/td>\n | 79.0<\/td>\n | 49.3<\/td>\n | 59.2<\/td>\n | —<\/td>\n | 63.8<\/td>\n<\/tr>\n |
\n| Agent<\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n<\/tr>\n |
\n| Long-horizon office work CoWorkBench<\/td>\n | 70.7<\/td>\n | 61.0<\/td>\n | 65.1<\/td>\n | —<\/td>\n | 68.2<\/td>\n<\/tr>\n |
\n| Professional job tasks JobBench<\/td>\n | 33.4<\/td>\n | 21.8<\/td>\n | 27.6<\/td>\n | —<\/td>\n | —<\/td>\n<\/tr>\n |
\n| Frontier agentic tasks Agents’ Last Exam<\/td>\n | Pass@1 20.4 Score 42.9<\/td>\n | Pass@1 10.6 Score 27.3<\/td>\n | Pass@1 13.2 Score 33.6<\/td>\n | —<\/td>\n | —<\/td>\n<\/tr>\n |
\n| General<\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n | <\/td>\n<\/tr>\n |
\n| Instruction following IFBench<\/td>\n | 79.5<\/td>\n | 69.1<\/td>\n | 79.1<\/td>\n | 77.0<\/td>\n | 62.5<\/td>\n<\/tr>\n |
\n| Scientific reasoning GPQA Diamond<\/td>\n | 89.2<\/td>\n | 87.8<\/td>\n | 90.3<\/td>\n | 83.5<\/td>\n | 91.3<\/td>\n<\/tr>\n |
\n| Multidisciplinary reasoning HLE<\/td>\n | 30.8<\/td>\n | 24.0<\/td>\n | 34.7<\/td>\n | 22.0<\/td>\n | 40.0<\/td>\n<\/tr>\n |
\n| Competitive coding LiveCodeBench v6<\/td>\n | 90.3<\/td>\n | 83.9<\/td>\n | 89.6<\/td>\n | —<\/td>\n | 88.8<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n These figures show that while maintaining the base model’s excellent knowledge, reasoning capabilities, and high potential in code generation and agent tasks, Signal achieves significant speedups and improved token efficiency by stripping away unnecessary preambles and redundant reasoning steps.<\/p>\n Strengths and Use Cases<\/h2>\nSignal is a fine-tuned model of Qwen3.8-27B designed to achieve lower generation latency and higher token efficiency. It cuts down on token usage without sacrificing answer quality by providing direct answers and omitting unnecessary preambles, excessive formatting, greetings, and explanatory narration. Furthermore, even in thinking mode, it reduces the tokens used to explain the process while preserving useful reasoning steps. It also retains the image input capabilities (vision encoder and projector) inherited from the base model.<\/p>\n <\/p>\n Hardware Requirements<\/h2>\nEstimated requirements (calculated by Local Model Watch)<\/strong> \u2014 27.8B parameters (taken from the base model Qwen\/Qwen3.8-27B)<\/p>\n |
![\"Reading](\"https:\/\/localmodelwatch.tsuchitsuchi.com\/wp-content\/uploads\/2026\/09\/lmw-lab-vision-chart-v1.png\")
<\/p>\n![\"Reading](\"https:\/\/localmodelwatch.tsuchitsuchi.com\/wp-content\/uploads\/2026\/09\/lmw-lab-vision-receipt-v1.png\")
<\/p>\n