Uploaded September 2026 | Updated November 2026, 4 days ago
On a 128K token document, Qwen 3.8 took 75 seconds to answer. Ornith took just 25. Why?
LLM generation speed is only half the story. How fast your model reads the prompt also matters. This is known as prefill.
▸ Ornith 1.5 9B reads around 10,000 tokens per second, so a 10,000 token prompt is processed in 1 second
▸ Qwen 3.8 27B reads 3,000 tokens per second, so the same 10,000 token prompt takes 3 seconds
▸ On a 128K token document, Qwen spends much more time just reading. That's why it was 3 times slower
Overall latency depends on both: reading speed and writing speed.
What do you check first, reading or writing speed? Tell me in the comments 👇
Follow for more AI benchmarks.
Full benchmark: kgptalkie.com/tutorials/llm-benchmarking/ornith-1-5-9b-vs-35b-a3b-benchmark
#LLM #AI #Shorts #Prefill #Qwen #LocalLLM #AIBenchmark
On a 128K token document, Qwen 3.8 took 75 seconds to answer. Ornith took just 25. Why?
LLM generation speed is only half the story. How fast your model reads the prompt also matters. This is known as prefill.
▸ Ornith 1.5 9B reads around 10,000 tokens per second, so a 10,000 token prompt is processed in 1 second
▸ Qwen 3.8 27B reads 3,000 tokens per second, so the same 10,000 token prompt takes 3 seconds
▸ On a 128K token document, Qwen spends much more time just reading. That's why it was 3 times slower
Overall latency depends on both: reading speed and writing speed.
What do you check first, reading or writing speed? Tell me in the comments 👇
Follow for more AI benchmarks.
Full benchmark: kgptalkie.com/tutorials/llm-benchmarking/ornith-1-5-9b-vs-35b-a3b-benchmark
#LLM #AI #Shorts #Prefill #Qwen #LocalLLM #AIBenchmark










