Launched this week
MiniMax M3.1 Flash
MiniMax’s Flash model, 3x faster and 70% cheaper for agents
270 followers
MiniMax’s Flash model, 3x faster and 70% cheaper for agents
270 followers
MiniMax M3.1 Flash is MiniMax’s new Flash model for agentic tasks, coding, computer use and cybersecurity. MiniMax says it is about 3x faster online and 70% cheaper for agentic tasks than its predecessor, with output at 200+ tokens per second. Pricing is $0.15 input and $0.60 output per 1M tokens. It’s on the MiniMax API, M Plan and MiniMax Code, and weights are promised later.
Free Options
Launch tags:Artificial Intelligence
Launch Team
Render: Get up to $100K in startup creditsThe fastest way to deploy and scale apps and agents
Promoted
MiniMax M3.1 Flash is MiniMax’s next-generation Flash model, built for agentic tasks, coding, computer use and cybersecurity. MiniMax calls it a “generational leap” over its predecessor.
An early version ran anonymously as “Space Bunny”, and MiniMax says it processed over 200 trillion tokens before the official launch.
What’s new in M3.1 Flash:
About 3x faster online inference than its predecessor, with more aggressive sparsity improving inference efficiency
A 70% lower usage cost for agentic tasks
Output at 200+ tokens per second in everyday use, with a median of 170+
Stronger agentic, coding, computer use and cybersecurity performance
Some numbers:
DeepSWE v1.1: 73.7%
OSWorld V2: 61.7 and ALE: 29.6, which MiniMax says rank first among Chinese models
BankerToolBench: 76.2
CyberGym: 82%, assessed by MiniMax’s internal security experts
Pricing is flat at $0.15 per 1M input tokens, $0.60 per 1M output tokens and $0.015 per 1M cached input tokens.
You can use it through the MiniMax API, the M Plan subscription and the MiniMax Code desktop app. Good fit if you’re building coding or computer-use agents and want a fast, low-cost model.
P.S. I hunt the latest and greatest launches in tech, SaaS and AI, follow to be notified → @rohanrecommends
As someone who always uses up my weekly tokens, a cheaper more efficient model is a godsend.
The DeepSWE and OSWorld numbers are what I would look at first for coding work, more than raw tokens per second. For long agent loops, the question I keep coming back to is how it holds up when a task needs several tool calls and a correction mid-run — does the speed hold, or does accuracy drop off first?
the price makes this interesting for longer agent runs
i would compare cost per finished task alongside tokens per second since retries can change the bill quite a bit
Three times faster at 70 percent cheaper is the kind of change that quietly makes a whole agent workflow viable, and 200 tokens per second should make those long tool loops feel less like waiting.