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?
| Dimension | MiniMax M2 | Zhipu GLM-5 | Kimi K2.5 | DeepSeek V4 |
|---|---|---|---|---|
| Release | 2025 Q4 | 2026 Q1 | 2026 Q2 | 2026 Q2 |
| Parameters | ~300B | Trillion MoE | Trillion MoE | Trillion MoE |
| LiveBench | ~78 | ~85 | ~87 | ~84 |
| Coding | Above average | #1 in China | #1 in China | #1 in China |
| Agent | Average | Strong (Swarm) | Strong (Multi-Agent) | Strong (Agent Integration) |
| Stock | -60% | Not listed | Not listed | Not 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:
- Trillion MoE architecture: Matching Kimi/DeepSeek scale, significantly increasing parameter count
- Agent-native capability: Multi-agent collaboration, tool calling, long-term memory
- Coding breakthrough: Competing with GLM-5 on SWE-Bench
- 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