 | MIT license and fully open source. MiMo-V2.5-Pro was just 3 points from Opus 4.7 max and the normal V2.5 is only a step behind SOTA. But both produce 75% and 68% non-hallucination rate. Best intel/hallucination model yet. V2.5 FP8 is like 316GB, you *might* be able to run a tight 3 bit quant with 128gb m5 max. From Gemma to Qwen3.6 to Kimi2.6 to Deepseek v4 to MiMo2.5, this probably is the best April. [link] [comments] |
