MiniMax M3.1 Flash
MiniMax has launched M3.1 Flash, a new AI model promising a generational leap in agentic tasks, coding, and cybersecurity, all while being three times faster and 70% cheaper. Formerly known as 'Space Bunny,' this model enters a competitive arena where its claimed performance and cost-effectiveness are being rigorously scrutinized against other 'Flash' models. HN users are debating whether it truly moves the Pareto frontier or if it's just another contender in an increasingly crowded and expensive AI landscape.
The Lowdown
MiniMax Research has officially unveiled M3.1 Flash, a new AI model positioned to redefine the balance of performance and cost for various complex tasks. This model, previously operating anonymously as 'Space Bunny,' boasts significant improvements in efficiency and capability across a range of benchmarks.
- Enhanced Agentic Capabilities: M3.1 Flash excels in complex enterprise workflows, scoring high on AutomationBench-AA and BankerToolBench for multi-step planning, financial compliance, and chained tool calls.
- Software Development Prowess: It achieves 73.7% on DeepSWE v1.1, demonstrating strong capabilities in understanding code, developing features, fixing issues, and validating work.
- Advanced Computer Use: The model scores well on OSWorld V2 and Agents’ Last Exam (ALE), indicating proficiency in interacting with real environments and executing long sequences of cross-application tasks.
- Cybersecurity Expertise: With an 82% score on CyberGym, M3.1 Flash shows leading capabilities in vulnerability discovery, code auditing, penetration testing, and malware analysis.
- Efficiency and Cost-Effectiveness: MiniMax claims inference speeds exceeding 200 TPS and a 70% reduction in usage costs for agentic tasks compared to its predecessor. Pricing is set at $0.15 input / $0.60 output / $0.015 cached input per 1M tokens.
- Availability: M3.1 Flash is accessible via the MiniMax API, M Plan, and MiniMax Code, with model weights to be released at a later date.
Overall, MiniMax M3.1 Flash aims to deliver high-level performance at a significantly lower cost, setting a new competitive standard for AI models, particularly in agentic applications. It represents MiniMax's continued effort to balance efficiency and intelligence in the evolving AI landscape.
The Gossip
Frontier Face-Off: Flash vs. Flash
Users are quick to compare MiniMax M3.1 Flash with its established competitor, DeepSeek V4.1 Flash. While MiniMax claims performance parity at lower cost, many commenters express skepticism, citing previous experiences with 'Space Bunny' or demanding significantly lower pricing to justify a switch. The discussion questions whether MiniMax truly pushes the performance-price frontier beyond what DeepSeek already offers, with some users stating it's 'already dead' if not considerably cheaper than competitors.
Benchmark Backlash & Real-World Reality
A recurring sentiment is skepticism towards benchmark scores, with users preferring real-world testing over advertised numbers. Some recall 'Space Bunny' (an earlier version of M3.1 Flash) making 'egregious errors,' suggesting a potential disconnect between advertised benchmark performance and practical utility. The importance of 'wall clock time' (actual execution speed) over mere cost savings is also highlighted for professional use cases, as benchmarks often omit this crucial metric.
The 'Flash' Factor: Fast, Cheap, or Just Big?
Commenters debate the meaning of 'Flash' in the context of these new AI models. While MiniMax advertises faster inference and lower costs, its total parameter count (428B) leads to questions about whether these models are truly 'small or light.' The consensus leans towards 'Flash' primarily referring to Mixture of Experts (MoE) architectures and general cost-efficiency, rather than small model size, as even 'Flash' models are becoming quite substantial.
Shadowban Shenanigans: HN's Content Filtering
A notable portion of the discussion revolved around the fact that MiniMax's domain was allegedly 'shadowbanned' on Hacker News, leading to multiple duplicate submissions of the story. Users speculated on the reasons behind the shadowban and called for moderation intervention to unshadowban the domain, acknowledging MiniMax's newsworthiness, highlighting a meta-discussion about HN's content submission policies.