AnalysisAI Models
14 hours ago
Worse quality with MTP on Qwen 3.6 and Gemma 4
User reports 80% of test cases show better output without MTP for Qwen 3.6 27B and Gemma 4. Tested with llama.cpp on 4x5070ti via OpenCode CLI.
·
14 hours ago
User reports 80% of test cases show better output without MTP for Qwen 3.6 27B and Gemma 4. Tested with llama.cpp on 4x5070ti via OpenCode CLI.