REDDIT · REDDIT// 34d agoBENCHMARK RESULT

MiniMax M2.5 still tops Qwen3-Coder-Next on 96GB rigs

In a LocalLLaMA field report, a user running 4 x 3090 Ti cards says MiniMax M2.5 still beats Qwen3-Coder-Next on greenfield coding work, even though QCN runs much faster. The complaint is not speed but planning depth: QCN handled a long PRD prompt quickly, yet still felt weaker than M2.5 when asked to turn that spec into useful project work through OpenCode.

// ANALYSIS

This is the local-coding-model tradeoff in one post: throughput gets attention, but greenfield software work still rewards models that can decompose specs and make better architectural calls.

–The poster reports QCN chewing through a roughly 48k-token prompt at solid speed, but still delivering output that felt shallow compared with MiniMax M2.5
–That lines up with MiniMax's official M2.5 positioning around spec writing, agentic planning, OpenCode compatibility, and strong SWE-Bench-style coding performance
–For self-hosters with 96GB of VRAM, the practical takeaway is that slower models can still be the better daily driver if they need less babysitting on full-project tasks
–It is still one anecdotal benchmark, not a controlled eval, but these rig-level reports often surface failure modes long before polished leaderboards do

// TAGS

minimax-m2-5llmai-codingreasoningbenchmarkself-hosted

DISCOVERED

34d ago

2026-03-09

PUBLISHED

34d ago

2026-03-09

RELEVANCE

8/ 10

AUTHOR

Ok-Measurement-1575