MiniMax open-sources the M3.1 Flash model, reducing usage cost by 70%
MiniMax announced on October 23 the open-sourcing of the M3.1 Flash model, which processed over 100 trillion tokens within 72 hours of launch. Compared with the previous generation M3, M3.1 Flash has comprehensively improved capabilities, with usage cost reduced by 70% and online inference speed increased by nearly 3 times. MiniMax stated it will work with chip manufacturers such as Huawei Ascend, NVIDIA, and AMD, as well as inference frameworks and communities such as vLLM, SGLang, and Hugging Face, to advance day-one open-source integration.
Why read this
AI score 73
The original text presents the cost and inference speed changes of the open-source model, and compares the capability focus of domestic Flash models in the same period.
Topics
Body · AI translation
Download the app
+1MiniMax open-sources M3.1 Flash, why have "Flash" models become a new AI trend? The Paper reporter Fan Jialai 2026-10-23 13:14 Source: The Paper ∙ 10% Company > Listen to full text Font size
Chinese large models are at it again — after DeepSeek officially announced its update, large-model company MiniMax followed suit with an update announcement.
On October 23, The Paper reporter learned that MiniMax (Xiyu Technology) announced the open-sourcing of the M3.1 Flash model, which processed over 100 trillion tokens within 72 hours of launch.
Compared with the previous-generation M3 model, the M3.1 Flash model achieves comprehensive capability improvements while cutting usage costs by 70%. Through optimizations including full-model sparse attention, quantization-aware training, and Power-aware scheduling, the model's online inference speed has increased by nearly 3x.
MiniMax stated that it will work with domestic and international chip vendors such as Huawei Ascend, T-Head, Kunlunxin, Moore Threads, NVIDIA, and AMD, as well as inference and cloud service platforms such as Together AI, Fireworks, and Nebius, to jointly advance day-one access for M3.1 Flash's open-source release. In addition, inference frameworks and open-source communities such as vLLM, SGLang, BAAI FlagOS, Hugging Face, and ModelScope will also work with MiniMax to promote the model's application across different chips, cloud platforms, and development environments, continuously expanding the global AI open-source ecosystem.
After M3.1 Flash's release, several international investment banks offered assessments. Citi believes the model consolidates MiniMax's competitive position and may accelerate its B2B commercialization progress in the developer and enterprise ecosystem. Citi noted that the focus of China's AI competition is accelerating toward end-to-end agent workflows and inference cost efficiency. Jefferies affirmed the model's capability improvements and cost-effectiveness in Agent tasks, coding, computer operation, and cybersecurity, maintaining a "Buy" rating on MiniMax.
Morgan Stanley noted that M3.1 Flash's pricing is highly competitive among mainstream Flash models in China, with significant advantages in Agent task costs as well. The bank believes that the combined improvements in intelligence level, throughput, and cost will help expand enterprise adoption, and maintains MiniMax's "Overweight" rating and HK$900 target price.
It is worth noting that Flash models have now become a "trend" in domestic large-model updates. Two days earlier, DeepSeek updated V4.1 Flash-1017 on October 17, focusing on enhancing full project development capabilities and improving task execution efficiency by shortening output length and task runtime.
Flash models are approaching the role of mainstay in complex tasks. Judging from public benchmark test results, domestic Flash models are already able to compete with leading closed-source models in coding and Agent evaluations while maintaining low invocation prices.
For developers, Flash models mean handling complex work at lower cost. As Flash models' task capabilities improve, model competition further extends to the delivery efficiency of complete tasks: not only completing the work, but also reducing retries, rework, and manual intervention. Only simultaneous improvements in capability, speed, and cost can drive more tasks into the high-frequency usage stage.
In this competitive direction, according to MiniMax, DeepSeek and MiniMax show advantages with different emphases. According to public comparison results, DeepSeek V4.1 Flash-1017 scores higher in software engineering evaluations, while M3.1 Flash performs more prominently in areas such as cross-SaaS business workflows and cybersecurity. The two can serve as complementary options for developers and enterprises to choose from based on the task, and also form a set of mutually referential, continuously competing product combinations in the domestic Flash field.
On October 22, MiniMax's Hong Kong stock briefly fell below its issue price, closing at around HK$160 (issue price HK$165). As of press time on the 23rd, MiniMax was up more than 6%. Responsible editor: Liu Xiuhao Image editor: Shen Ke Proofreader: Zhang Yan The Paper news tip: 021-962866 The Paper, unauthorized reproduction prohibited +1 Favorite I want to report #MiniMaxOpenSourcesM3.1 #FlashModel View more