header-langage
简体中文
繁體中文
English
Tiếng Việt
한국어
日本語
ภาษาไทย
Türkçe
Scan to Download the APP
Partnered With:

MiniMax M3 Officially Joins DeepSWE's Official Ranking: Approval Rate Rises to 20.4%, Surpassing GLM and Xiaomi Mimo

According to TrendForce Beating Monitor, a benchmark platform focusing on evaluating the ability of large models to tackle long-range software engineering tasks, the latest ranking update of DeepSWE, an official benchmark, has included the MiniMax encoding model MiniMax M3. In the evaluation of 113 real-world open-source tasks, MiniMax M3 achieved a Pass@1 rate of 20.4%. Although this performance still lags behind Kimi-k2.6 from the DarkSide of the Moon (23.9%), MiniMax M3 has successfully surpassed Xiaomi's mimo-v2.5-pro (19.5%), Zhiku's GLM-5.1 (17.5%), Google's Gemini 3.1 Pro (9.7%), and DeepSeek V4 Pro (7.5%).

Compared to the independent evaluation results released by the developer @bleysg on June 2nd, MiniMax M3 has shown significant improvement in performance. In the prior independent test, M3 had a strict pass rate of only 13.3% (15/113) within a 90-minute regular time limit, trailing behind Kimi and GLM in the reported model performance. The 20.4% pass rate in the official test marks a significant enhancement, mainly attributed to a major inference upgrade deployed by MiniMax on June 3rd. This upgrade fixed the issue of occasional generation of abnormal tokens and optimized the long-context cache efficiency, leading to a notable improvement in the model's code accuracy in long-range programming and closing the gap with leading models. It is worth noting that in terms of Pass@4 (pass rate allowing 4 attempts), MiniMax M3 achieved 48.7%, tied with Kimi-k2.6, significantly higher than GLM-5.1 at 38.9%.

However, the high resource and computational costs remain a weakness for MiniMax M3. In the official benchmark test, M3 has a median of 311 steps in resolving a single task for the agent, a median output token of 91,000, and a median single-task average cost of $5.04. In comparison, Kimi-k2.6 has a median step count of only 136.5 to resolve a single task with a median cost of $2.70; while Xiaomi's mimo-v2.5-pro has a median step count of only 114.5 and a lower median cost of $1.68. The data indicates that although M3 performs well in terms of problem-solving upper limits and pass rates, there is still significant room for improvement compared to other mainstream models domestically in terms of computational efficiency and inference cost optimization.

Correction/Report
Submit
Add Library
Visible to myself only
Public
Save
Choose Library
Add Library
Cancel
Finish