MiniMax M2.7 draws mixed local reports

Hot take: this looks less like a simple “bad model” report and more like a stability/polish problem surfacing under realistic coding settings, especially at the recommended high-entropy sampling regime.

• –The report is specifically about raw HF weights on vLLM, so this is useful signal for self-hosters, not just API users.
• –The biggest complaint is inconsistency: the same evals that worked reliably on M2.5 are now producing more variable results.
• –The formatting issues, spacing regressions, and stray Chinese characters point to output hygiene problems in addition to task quality.
• –The thread suggests M2.7 may need tighter decoding for code workflows than M2.5 did, despite MiniMax’s recommended defaults.
• –This is still anecdotal, but it matches a common pattern where frontier models need more careful sampling control to behave predictably in production.