The domestic big model open source camp has once again played a tough role. On April 10th, Xiyu Technology fulfilled its promise and officially opened sourced the Minimax 2.7 model released in March. This is another domestically produced open-source model that has reached the international top level in programming ability, following the GLM-5.1 of Zhipu.
Programming ability: This time it's really close to the ceiling
The most noteworthy aspect of Minimax 2.7 is the improvement of programming ability. Several key data:
What does the SWE Pro benchmark score of 56.22% mean? Claude Opus is currently the top benchmark for programming ability, and Minimax 2.7 is almost on the same level as it. For an open-source model, this achievement is already quite high.
MMClaw evaluation shows that in OpenClaw environment, the actual performance of Minimax 2.7 is significantly better than the previous generation M2.5 and close to the latest Sonnet 4.6. The evaluation in OpenClaw environment has more reference value - this is the performance in real application scenarios, not the numbers on the leaderboard.
More importantly, there is a cost advantage. Compared to overseas charging models such as GPT-5.4 Pro and Claude Opus, Minimax 2.7 provides developers with a low-cost and high-performance option. A team with a large amount of programming tasks and limited budget will have a significant difference.
The open source landscape is reshaping
The open-source release of Minimax 2.7 has once again changed the landscape of the global open-source big model market.
At present, domestic open-source models have occupied an important share. GLM-5.1 and Minimax 2.7 have been open sourced successively, which means that developers have two more reliable options when choosing open source models with strong programming capabilities and controllable costs.
A noteworthy signal is that, according to reports, with strategic adjustments, the future Qwen3.6 Plus may no longer use a completely open-source operating model. If this rumor comes true, it will be a loss for the open source ecosystem, but for players who continue to open source, it will actually mean a redistribution of developer resources.
DeepSeek V4: The Next Suspense of This Competition
After other major companies fell, the biggest suspense in the market returned to DeepSeek. Liang Wenfeng has stated that V4 will be released in late April, which will natively support multimodality and provide three modes of "fast, expert, and visual", and fully adapt to domestic AI hardware systems.
There is still a gap between domestic large-scale models and OpenAI in terms of computing power investment scale, but through high-frequency iteration and differentiated competition, the generation gap is indeed rapidly narrowing in AI programming and specific industry applications. The open-source release of Minimax 2.7 is a microcosm of this trend.
