MiniMax M2.7 vs GLM-5.1 vs Kimi K2.5 — Chinese Frontier Models Compared
Chinese AI labs are producing some of the most capable and affordable coding models in 2026. Three stand out: MiniMax M2.7 from Shanghai, GLM-5.1 from Z.ai in Beijing, and Kimi K2.5 from Moonshot AI.
Each takes a different approach, and each has a distinct sweet spot. This comparison breaks down where each model excels and which one fits your workflow.
Update (April 21, 2026): Moonshot AI released Kimi K2.6, which scores 80.2% on SWE-Bench Verified and scales Agent Swarm to 300 sub-agents. This significantly strengthens Kimi’s position in this comparison. See GLM 5.1 vs Kimi K2.6 for the updated head-to-head.
For a broader view of how these compare to Western alternatives, see our AI model comparison.
Specifications compared
| MiniMax M2.7 | GLM-5.1 | Kimi K2.5 | |
|---|---|---|---|
| Company | MiniMax (Shanghai) | Z.ai (Beijing) | Moonshot AI (Beijing) |
| Total parameters | 230B | 754B | 1T |
| Active parameters | 10B | 40B | 32B |
| SWE-bench Pro | 56.22% | 58.4% | — |
| Input price | $0.30/1M | $1.00/1M | $0.60/1M |
| Speed | 100 tok/s | 55 tok/s | — |
| Context window | 200K | 200K | 256K |
| License | Open weights | MIT | MIT |
| Unique feature | Self-evolving | 8hr autonomous | Agent Swarm |
All three use Mixture-of-Experts architectures, meaning only a fraction of total parameters activate per token. This keeps inference costs low while maintaining large knowledge bases.
The active parameter counts range from 10B for MiniMax to 40B for GLM, directly affecting reasoning depth per forward pass.
MiniMax M2.7 — speed and value champion
MiniMax M2.7 is the fastest and cheapest of the three. At 100 tokens per second and $0.30 per million input tokens, it delivers remarkable throughput for its price.
The self-evolving capability means the model can improve its own responses through iterative refinement. This is useful for code generation where the first attempt may need adjustment.
With only 10B active parameters, MiniMax trades some reasoning depth for speed. It handles routine coding tasks well — boilerplate generation, simple refactoring, documentation writing, standard CRUD operations.
Where it falls short is complex multi-step reasoning requiring deeper analysis.
For teams processing high volumes of straightforward coding requests, MiniMax offers the best cost-per-token value of any frontier-class model.
GLM-5.1 — raw coding quality leader
GLM-5.1 leads the group on SWE-bench Pro at 58.4%. Its standout feature is the ability to work autonomously for up to 8 hours on a single task.
You can hand GLM a complex feature request and come back to find it completed, tested, and documented. The model handles planning, implementation, testing, and documentation as a continuous workflow.
The 40B active parameters give GLM the deepest reasoning capability of the three. It handles complex architectural decisions, subtle bug detection, and nuanced code review better than MiniMax or Kimi.
The trade-off is speed — at 55 tokens per second, it is roughly half as fast as MiniMax.
GLM was trained on Huawei Ascend chips rather than NVIDIA GPUs. Despite the different training hardware, performance is competitive with NVIDIA-trained models.
The Z.ai Coding Plan at $18 per month provides API access compatible with Claude Code, making it one of the most affordable ways to get frontier-class coding assistance.
Kimi K2.5 — parallel execution specialist
Kimi K2.5 differentiates itself with Agent Swarm, enabling parallel execution of multiple tasks simultaneously.
Instead of processing coding tasks sequentially, Kimi spins up multiple agent instances working on different parts of a problem at the same time.
This makes Kimi effective for:
- Batch refactoring across large codebases
- Migrating projects between frameworks
- Running parallel test generation
- Any workflow where tasks decompose into independent subtasks
The 256K context window is the largest of the three, helping when working with extensive codebases.
At $0.60 per million input tokens, Kimi sits between MiniMax and GLM on price. The 32B active parameters provide a good balance between reasoning depth and speed.
Choosing the right model
Pick MiniMax M2.7 if cost is your primary concern and coding tasks are relatively straightforward. At $0.30/M with 100 tok/s throughput, nothing matches its value for routine work.
Pick GLM-5.1 if you need the highest coding quality and can tolerate slower speeds. The 8-hour autonomous capability and 58.4% SWE-bench Pro score make it best for complex software engineering.
Pick Kimi K2.5 if your workflow benefits from parallel execution. Agent Swarm is uniquely powerful for batch operations across large codebases.
For many teams, the best approach is using multiple models. Route simple tasks to MiniMax for speed, send complex problems to GLM for quality, and use Kimi when parallelization speeds up large-scale operations.
The bigger picture
Chinese AI labs now produce four of the top ten coding models globally. A year ago, the conversation was Claude versus GPT. Now developers choose from six or more frontier-class models.
This competition benefits everyone. MiniMax at $0.30/M, DeepSeek at $0.27/M, and GLM’s Coding Plan at $18/month mean frontier-class AI coding is accessible to individual developers, not just well-funded enterprises.
FAQ
Which Chinese AI model is best?
It depends on priorities. GLM-5.1 has the highest coding benchmarks (58.4% SWE-bench Pro) and best autonomous capabilities. MiniMax M2.7 is fastest and cheapest. Kimi K2.5 offers unique parallel execution through Agent Swarm. For raw coding quality, GLM leads. For value, MiniMax wins. For parallel workflows, Kimi is best.
Are these models free?
None are completely free via API, but all are very affordable. MiniMax M2.7 starts at $0.30 per million input tokens. GLM-5.1 offers a Coding Plan at $18/month. Kimi K2.5 costs $0.60/M input tokens. All three offer free tiers or trial credits for new users.
Can I use them with Claude Code?
GLM-5.1 works directly with Claude Code through the Z.ai Coding Plan, which provides a compatible API endpoint. MiniMax M2.7 and Kimi K2.5 can be routed through OpenRouter or similar API aggregators to work with Claude Code and other tools supporting OpenAI-compatible endpoints. Kimi also has its own CLI tool.
How do they compare to GPT-5?
On coding benchmarks, GLM-5.1 and Kimi K2.5 are competitive with GPT-5 on many tasks while costing a fraction of the price. MiniMax M2.7 falls slightly behind on complex reasoning but matches GPT-5 on routine coding. The main advantage is price — you get 80-90% of GPT-5’s coding performance at 10-30% of the cost.