MiniMax M3.1 Flash Arrives on SambaCloud at Over 200 Tokens per Second

What MiniMax M3.1 Flash is built for

MiniMax M3.1 Flash is the latest Flash model from MiniMax, designed for agentic tasks, coding, computer use and cybersecurity. SambaNova describes it as a generational leap in capability over the previous model, while making inference nearly 3x faster.

The speed and cost gains come from more aggressive sparsity than in its predecessor. According to the post, this lowers overall usage cost for agentic tasks by 70% and makes online inference nearly 3x faster. On SambaCloud, SambaNova says this efficiency combines with its dataflow architecture to reach over 200 tokens/s.

Customers who currently call MiniMax M3 on SambaCloud should note a change: those requests are being rerouted to MiniMax M3.1 Flash. SambaNova asks users to update their model string to MiniMax-M3.1-Flash.

Benchmark results reported

The figures below are the benchmark results given in the SambaNova post, which draws on MiniMax's announcement.

• Coding: 73.7% on DeepSWE v1.1, a test where the model must understand a codebase, build features, fix issues and validate its own work through testing.
• Computer use: 61.7 on OSWorld V2 and 29.6 on Agents' Last Exam (ALE). The post says the model ranks first among Chinese models on both.
• Cybersecurity: 82% on CyberGym, which the post links to vulnerability discovery and validation.
• Finance workflows: 76.2 on BankerToolBench, which stresses multi-step planning, consistency across documents and files, financial compliance and chained tool calls.

The post also says the model ranks among the top performers on AutomationBench-AA, completing enterprise workflows across SaaS systems in areas such as sales, marketing, operations, customer service, finance and HR. In MiniMax's internal evaluations by security experts, the model performed strongly in code auditing, vulnerability patching, penetration testing, malware analysis and reverse engineering.

Pricing and getting started

SambaNova lists the following pricing for MiniMax-M3.1-Flash on SambaCloud: $0.03 for cached input tokens, $0.30 per 1M input tokens and $1.20 per 1M output tokens.

The post includes a short Python example using the SambaNova client. It sends a coding prompt, asks the model to critique its own solution and then produce an improved version. The example sets temperature to 1.0 and top_p to 0.95. Users can also try the model in the SambaCloud playground or generate an API key to integrate it into an application.

What it means for agent builders

SambaNova frames the release around production agents, where each step in a long loop is a model call. In that setting, it argues, speed and low task cost matter most. The company says that across AutomationBench, DeepSWE and ALE, the model delivers performance on par with top models at a very low task cost.

These performance claims come from SambaNova and MiniMax rather than independent testing. For teams already using MiniMax M3 on SambaCloud, the practical step is to switch to the new model string, since traffic is already being served by M3.1 Flash.