Community · Oct 23, 2026, 18:22 UTC

Benchmark ranks Basalt and Infernix ahead of other Qwen 3.8 FN engines on a single RTX 5090

Basalt led decode speed in most tests, reaching 354 tokens per second on short structured output, and had the highest prefill speed at 50K and above. The results come from the engine developer, who is one of the benchmarked authors, and llama.cpp ran with MTP off.

• ModelQwen 3.8 FN
• Topicbenchmark
• Topicopen-weights
• Topicchips