Two major MoE models dropped this week — Moonshot's Kimi K2.7 Code and MiniMax's M3. Both target agentic coding workflows, but take starkly different architectural approaches. Here's how they stack up.
Kimi K2.7 Code: Coding-First Agentic Powerhouse
Moonshot AI built K2.7 Code on top of K2.6 with a singular focus: real-world long-horizon coding tasks. The improvements are substantial — 30% reduction in thinking-token usage compared to K2.6, meaning faster iteration loops and lower API costs for agentic workflows.
Architecture: Mixture-of-Experts with 1T total parameters, 32B activated per token. 61 layers (1 dense + 60 MoE), 384 experts with 8 selected per token, 1 shared expert. Uses Multi-head Latent Attention (MLA) with 7168 hidden dim and 64 attention heads. 256K context window. Includes MoonViT vision encoder (400M params) for multimodal input.
Key benchmarks (vs K2.6): - Kimi Code Bench v2: 50.9 → 62.0 (+21.8%) - Program Bench: 48.3 → 53.6 (+11%) - MLS Bench Lite: 26.7 → 35.1 (+31.5%) - MCP Atlas: 69.4 → 76.0 (+9.5%) - MCPMark Verified: 72.8 → 81.1 (+11.4%)
For context, GPT-5.5 scores 69.0 on Kimi Code Bench v2 and Claude Opus 4.8 scores 67.4. K2.7 Code still trails the frontier on pure coding, but its MCPMark Verified score (81.1) actually beats Opus 4.8 (76.4) — a strong signal for real-world MCP tool-use scenarios.
The model uses native INT4 quantization (same method as K2-Thinking), keeping deployment costs manageable despite the 1T parameter footprint.
MiniMax M3: Multimodal Efficiency Play
MiniMax took a different bet. Where K2.7 Code doubles down on coding, M3 goes wide — native multimodal training from step one, fusing text, image, and video. The headline feature is MiniMax Sparse Attention (MSA), which delivers 9× prefill and 15× decode speedups compared to M2 at 1M context, reducing per-token compute to 1/20th.
Architecture: ~428B total parameters, ~23B activated per token. Native multimodal. 1M context window — 4× longer than K2.7 Code. Three reasoning modes: enabled, adaptive (model decides), and disabled (minimize latency).
The sparse attention story is M3's real differentiator. Long-context inference has been a persistent bottleneck for production deployments. If MSA delivers as promised, M3 could be the model that makes million-token contexts economically viable for everyday applications rather than research demos.
MiniMax also published their technical report (arXiv:2606.13392) and open-sourced the MSA operator on GitHub.
Head-to-Head
| Dimension | Kimi K2.7 Code | MiniMax M3 |
|---|---|---|
| Total params | 1T | ~428B |
| Activated params | 32B | ~23B |
| Context length | 256K | 1M |
| Attention | MLA | MSA (sparse) |
| Multimodal | Vision (MoonViT) | Native (text+image+video) |
| Focus | Coding/agentic | Multimodal/cowork |
| Quantization | Native INT4 | Standard |
| Experts | 384 (8 active) | MoE |
What This Means for You
Choose K2.7 Code if: You're building coding agents, MCP-based tool workflows, or software engineering pipelines. The token-efficiency gains over K2.6 compound fast in long agentic loops, and the MCPMark scores are genuinely impressive.
Choose M3 if: You need long-context processing (document analysis, video understanding, large codebases), multimodal capabilities, or want to serve a single model across diverse tasks. The sparse attention efficiency story makes 1M contexts practical.
Both models are available now on Hugging Face and through their respective API platforms. Kimi K2.7 Code runs through platform.moonshot.ai, and M3 through platform.minimax.io.
The Bigger Picture
These releases signal where the MoE frontier is heading. Pure parameter count matters less than activation efficiency and specialized capability. K2.7 Code proves that focused post-training on a general MoE base can yield domain-specific gains that rival purpose-built models. M3 proves that sparse attention is mature enough for production million-token contexts.
Expect more of this pattern: large MoE backbones with targeted fine-tuning for specific agentic domains, paired with attention innovations that make long contexts economically viable.