ResearchBenchmarksAutoregressive Generation on Long-context evaluation document 40,022 tokensAutoregressive Generation on Long-context evaluation document 40,022 tokensFollow26.4Time To First Token (ms)Nectar13.576100.138186.7273.262May 10, 2026Evaluation resultsMethodMethodLinksTime To First Token (ms)Throughput (Tokens/s)Memory Usage (GB)NectarModel=Qwen3-1.7B, Infe...Model=Qwen3-1.7B, Inference Protocol=Nectar2026.0526.444.27.7NectarModel=Qwen3-4B, Infere...Model=Qwen3-4B, Inference Protocol=Nectar2026.0544.618.416.7BaseModel=Qwen3-1.7B, Infe...Model=Qwen3-1.7B, Inference Protocol=Base2026.051839.524.9BaseModel=Qwen3-4B, Infere...Model=Qwen3-4B, Inference Protocol=Base2026.053476.338.6