MiniMax-M2.7 开源模型性能媲美 GPT-5，推理速度 440+ tokens/s
开源模型首次在性能上追平顶级闭源模型，且推理速度和成本优势巨大，做 AI 应用开发或模型选型的团队值得立即体验。
开发者 Santiago 表示首次感到开源权重模型不可忽视，MiniMax-M2.7 以 230B 参数在 SambaNova 上实现 440+ tokens/s 的极速推理。
该模型在 SWE-Pro 上得分 56.22%，Terminal Bench 2 得分 57.
0%，SWE Multilingual 得分 76.5%，性能接近 Opus 4.
6 和 GPT-5.4 级别。
使用成本仅为专有模型的 5%，且完全开源。SambaNova 提供免费 playground 供测试。
For the first time, I feel open-weight models are impossible to ignore.
We are at a point where these models are competitive with the best models out there.
MiniMax-M2.7 is the latest beast to come out, and I'm running it at 440+ tokens/s.
230B parameters. It's a beast.
Just for comparison, I've found that Gemma4 31b is good enough for many of the things I do, so imagine what an additional 200B parameters bring here.
Of course, I can't run MiniMax-M2.7 locally, so I use SambaNova.
• Extremely fast inference (probably one of the fastest in the market) • Extremely cheap (around 5% of what you'd pay for proprietary models) • MiniMax 2.7 scores 56.22% on SWE-Pro, 57.0% on Terminal Bench 2, and 76.5% on SWE Multilingual
That puts it around the Opus 4.6/GPT-5.4 league, with the difference that MiniMax is open-weight.
I recorded a real-time video of MiniMax-M2.7 running on SambaNova, so you can get a sense of how fast it is. I'm not using streaming here. The model is consistently running at ~440 tokens/second.
Use this playground to test MiniMax M2.7: https://t.co/c2jUA8PsJh
Thanks to the team for partnering with me on this post.
This report is categorized under AITOP's "Models" desk, focusing on China's AI industry and domestic foundation model ecosystem. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include 开源模型, MiniMax-M2.7, 推理速度, SambaNova.
Source attribution: curated from Ate-a-Pi (@svpino).
