MiniMax's 428B M3.1 Flash cuts agent costs by 70% with 3x faster inference
The Chinese AI developer launches a 428-billion-parameter update that reportedly slashes agent execution expenses by 70 percent.
Key points
- MiniMax Group Inc. released M3.1 Flash, an update to its M3 large language model.
- The company claims the new model reduces AI agent running costs by 70 percent.
- Inference speeds are reported to be nearly three times faster than the previous version.
MiniMax Group Inc. released M3.1 Flash today, a faster and more economical iteration of the M3 large language model it introduced in June. According to SiliconANGLE, the Chinese artificial intelligence developer positioned this update specifically to address the high expenses associated with running autonomous AI agents.
The release targets the financial burden of agentic tasks, which frequently require dozens of separate model calls to complete a single objective. MiniMax stated that M3.1 Flash represents its ongoing exploration of balancing efficiency with capability. The company reported that agentic tasks now cost 70 percent less to execute than on the prior model.
Technical Specifications And Performance
The model retains a size of 428 billion parameters, similar to its predecessor. According to the developer guide cited by SiliconANGLE, the increased speed stems from more aggressive sparsity during inference. Reasoning capabilities remain active for every request, allowing developers to select from five effort levels ranging from Low to Max.
Input modalities include text, images, and video, with prompt lengths supported up to 1 million tokens. SiliconANGLE noted that inference speeds are close to three times as fast as the previous version, with output topping 200 tokens per second during everyday peak usage.
Benchmark Results And Market Position
In internal testing, M3.1 Flash achieved a score of 61.7 on OSWorld V2, a benchmark for computer operation skills. MiniMax claimed no Chinese model has surpassed this score. On BankerToolBench, which evaluates banking workflows, the model scored 76.2.
For security applications, the model scored 82 percent on CyberGym. SiliconANGLE reported that University of California, Berkeley researchers created this benchmark to test if AI agents can reproduce real software vulnerabilities. MiniMax listed code auditing, penetration testing, and malware analysis as potential use cases.
Pricing is set at 15 cents per million input tokens. This aligns directly with the V4.1-Flash model from Hangzhou DeepSeek Artificial Intelligence Basic Technology Research Co. Ltd., which charges the same rate off-peak. On the DeepSWE v1.1 software engineering test, DeepSeek’s model resolved
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe
- Sudden role changes in output signal a successful override of system instructions.
- Hidden metadata in images or documents often carries invisible injection payloads.
- Immediate isolation of the affected session prevents lateral movement within the AI workflow.
Prompt injections exploit the model’s ability to follow instructions by hiding malicious commands in input data. Isolate affected sessions immediately and implement external guardrails to filter inputs before they reach the model.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
