MiniMax 3.0 on the Horizon: M2 Falling Behind, Stock Under Pressure, The Life-or-Death Battle for China's Second-Tier AI Models

MiniMax 3.0 on the Horizon: M2 Falling Behind, Stock Under Pressure, The Life-or-Death Battle for China's Second-Tier AI Models

Core Conclusion

MiniMax is experiencing the most painful “falling behind anxiety” among Chinese AI model companies. Since M2’s release, it has been consistently surpassed by Zhipu GLM-5 and Moonshot Kimi K2.5 in public benchmarks like LiveBench and Arena, with stock price dropping over 60% from its peak. Community rumors suggest MiniMax 3.0 is imminent, but the window for this company—once considered a candidate for “China’s No.1 AI model”—is rapidly closing.

Data Comparison: What Is MiniMax Losing?

DimensionMiniMax M2Zhipu GLM-5Kimi K2.5DeepSeek V4
Release2025 Q42026 Q12026 Q22026 Q2
Parameters~300BTrillion MoETrillion MoETrillion MoE
LiveBench~78~85~87~84
CodingAbove average#1 in China#1 in China#1 in China
AgentAverageStrong (Swarm)Strong (Multi-Agent)Strong (Agent Integration)
Stock-60%Not listedNot listedNot listed

Key issue: MiniMax M2 uses MoE architecture with comparable parameter scale to competitors, but lags across the board in actual benchmarks. Community analysis suggests this is not just a model scale issue, but potentially a systemic gap in training data quality, engineering capability, and alignment strategy.

Competitive Landscape: China’s “One Super, Many Strong” Is Taking Shape

The current Chinese model landscape has clearly differentiated:

  • First Tier: Qwen 3.6 (open-source ecosystem leader), Kimi K2.5 (outstanding Agent capability), DeepSeek V4 Pro (best price-performance)
  • Second Tier: Zhipu GLM-5 (solid tech but weak commercialization), MiniMax (strong product but lagging model)
  • Chasers: Baichuan, SenseNova, Mimo, etc.

MiniMax’s unique advantage lies in product and commercialization capability. Its consumer-facing AI products (like Hailuo AI) deliver excellent user experience, but lagging in underlying model capability creates an awkward “good product + weak model” combination. Once competitors rapidly catch up at the product layer, MiniMax’s last moat will be filled.

MiniMax 3.0: Possible Comeback Paths

Community rumors about MiniMax 3.0, combined with industry trends, suggest 3.0 may focus on:

  1. Trillion MoE architecture: Matching Kimi/DeepSeek scale, significantly increasing parameter count
  2. Agent-native capability: Multi-agent collaboration, tool calling, long-term memory
  3. Coding breakthrough: Competing with GLM-5 on SWE-Bench
  4. Vertical scenario optimization: Deep optimization in MiniMax’s strong AI entertainment and social scenarios

But the problem is that the catch-up window is closing. Kimi has already previewed K3 (2026 Q3), GLM is iterating to GLM-5.1, and DeepSeek keeps cutting prices. MiniMax 3.0 not only needs to catch up but must establish new differentiation before competitors’ next iteration.

Action Recommendations

For developers:

  • Not recommended as primary model at this stage, unless specific scenario optimization is needed
  • Monitor 3.0’s first-week benchmark data after release to confirm capability leap magnitude
  • If already integrated with MiniMax, prepare 3.0 migration plan

For investors:

  • MiniMax stock has priced in most pessimistic expectations, but 3.0 performance will determine the inflection point
  • Watch user growth and API call volume changes after 3.0 release
  • Risk: If 3.0 still cannot catch up, MiniMax may face further valuation markdown