AI Skills AI技能 7d ago Updated 7d ago 更新于 7天前 52

Inside Kimi K3: How Moonshot AI Built the Largest Open-Source Model 深入 Kimi K3:月之暗面如何构建最大开源模型

Kimi K3 is the largest open-weight model as of mid-2026 at 2.8 trillion parameters, with Moonshot AI taking a novel approach to the memory bottleneck in long-context LLMs rather than simply scaling hardware. Kimi Delta Attention (KDA) replaces standard KV-cache-based attention with a fixed-size whiteboard-style memory that selectively forgets unimportant information using a learned Delta Rule with per-head forgetting dials, cutting memory use by ~75%. K3 combines KDA (69 of 93 attention layers) Kimi K3 是截至 2026 年中最大的开源权重模型,参数量达 2.8 万亿,成为首个进入 3 万亿参数级别的开源模型 核心创新是 Kimi Delta Attention(KDA)机制,通过"选择性遗忘"替代标准 Transformer 的 KV cache 线性增长模式,将内存使用降低约 75% 采用 KDA 与 Gated MLA 混合架构(约 3:1 比例),兼顾高效固定内存与精确细节召回,支持百万 token 上下文窗口 结合 Stable LatentMoE(896 专家中每 token 激活 16 个)、Attention Residuals 和量化感知训练等优化,实现约 2

72
Hot 热度
75
Quality 质量
78
Impact 影响力

Analysis 深度分析

TL;DR

  • Kimi K3 is the largest open-weight model as of mid-2026 at 2.8 trillion parameters, with Moonshot AI taking a novel approach to the memory bottleneck in long-context LLMs rather than simply scaling hardware.
  • Kimi Delta Attention (KDA) replaces standard KV-cache-based attention with a fixed-size whiteboard-style memory that selectively forgets unimportant information using a learned Delta Rule with per-head forgetting dials, cutting memory use by ~75%.
  • K3 combines KDA (69 of 93 attention layers) with Gated Multi-Head Latent Attention (24 layers) for exact recall, plus Attention Residuals, Stable LatentMoE with 896 experts (16 active per token), and several training optimizations including Quantile Balancing and Quantization-aware training in MXFP4/MXFP8.
  • The model supports a 1-million-token context window with native multimodal (text, image, video) understanding and achieved top scores on WebDev, Program Bench, SWE Marathon, BrowseComp, and OmniDocBench at launch, though it trails Claude Fable 5 and GPT-5.6 Sol on overall benchmarks.
  • KDA is the culmination of Moonshot's long-context research lineage (K1.5 → K2 → K2.5 → Kimi Linear), reflecting a strategic shift from brute-force compute scaling toward intelligent memory management in large models.

Why It Matters

Kimi K3 demonstrates that the next frontier in scaling large language models is not raw parameter count alone but intelligent memory architecture—teaching models to forget selectively rather than hoard everything. For AI practitioners and researchers, this signals a paradigm shift: as models approach trillion-parameter scales, the quadratic cost of standard attention makes architectural innovation in memory management essential for practical deployment. Moonshot's open-weight release and vLLM integration also lower the barrier for the community to experiment with and build upon these techniques.

Technical Details

  • Kimi Delta Attention (KDA): A recurrent-style attention mechanism with a fixed-size memory buffer. Each attention head has 128 independent forgetting dials governed by two learned parameters: β (writing strength) and α (forgetting factor). The "erase then write" Delta Rule prevents catastrophic interference from new information overwriting old memories, building on Gated DeltaNet.
  • Hybrid Attention Architecture: K3 uses 69 KDA layers interleaved with 24 Gated Multi-Head Latent Attention (Gated MLA) layers at roughly a 3:1 ratio. KDA handles efficient fixed-size memory for long contexts, while Gated MLA (an evolution of DeepSeek's MLA) compresses the KV cache into a latent representation to preserve exact token-by-token recall for critical details.
  • Stable LatentMoE: A Mixture-of-Experts design with 896 total experts, activating only 16 per token, yielding ~104 billion active parameters out of 2.8 trillion total. This sparsity enables massive scale while keeping per-token compute tractable.
  • Attention Residuals (AttnRes): A drop-in replacement for standard residual connections that mitigates PreNorm dilution by allowing later layers to selectively pull representations from earlier layers, improving training efficiency by ~25% for under 2% additional compute cost.
  • Training Optimizations: Quantile Balancing replaces auxiliary loss terms for expert load balancing by deriving routing biases directly from router-score quantiles. Per-Head Muon extends the Muon optimizer for independent attention-head optimization. A custom Sigmoid Tanh Unit (SiTU) activation function replaces GeLU/SwiGLU. Quantization-aware training from SFT onward uses MXFP4 weights with MXFP8 activations, eliminating post-hoc quantization loss.
  • Multimodal & Scale Specs: MoonViT-V2 vision encoder (~401M parameters) handles image and video input. Vocabulary size is ~160,000 tokens. Context window is 1 million tokens. Released as ~96 shards totaling ~1.56 TB under a custom Kimi K3 License.

Industry Insight

  • The success of KDA suggests that memory efficiency, not just scale, will be the defining competitive advantage in the next generation of open-weight models. Labs that solve the quadratic attention cost problem will unlock practical million-token contexts for trillion-parameter models, opening doors to applications (full-codebase reasoning, long-document analysis, extended agentic workflows) that are currently infeasible.
  • Moonshot's hybrid approach—combining efficient forgetting (KDA) with compressed exact recall (Gated MLA)—offers a practical blueprint for other labs: pure linear attention sacrifices too much fidelity, while pure standard attention is too expensive. The 3:1 interleaving ratio in K3 is a design choice worth studying and potentially adapting.
  • The open-weight release with vLLM integration and quantization-aware training in MXFP4/MXFP8 lowers deployment barriers significantly. This could accelerate community adoption of trillion-parameter models and spur further innovation in efficient attention mechanisms, particularly as hardware constraints continue to limit access to the largest proprietary models.

TL;DR

  • Kimi K3 是截至 2026 年中最大的开源权重模型,参数量达 2.8 万亿,成为首个进入 3 万亿参数级别的开源模型
  • 核心创新是 Kimi Delta Attention(KDA)机制,通过"选择性遗忘"替代标准 Transformer 的 KV cache 线性增长模式,将内存使用降低约 75%
  • 采用 KDA 与 Gated MLA 混合架构(约 3:1 比例),兼顾高效固定内存与精确细节召回,支持百万 token 上下文窗口
  • 结合 Stable LatentMoE(896 专家中每 token 激活 16 个)、Attention Residuals 和量化感知训练等优化,实现约 2.5 倍于 K2 的缩放效率
  • 在 WebDev、Program Bench、SWE Marathon 等编码与 Agent 基准上领先,但整体性能仍落后于 Claude Fable 5 和 GPT-5.6 Sol 等闭源旗舰模型

为什么值得看

Kimi K3 展示了开源模型突破参数规模与长上下文计算瓶颈的新路径,证明通过架构创新(而非单纯堆硬件)可实现万亿级参数的实用化部署。其选择性记忆机制为行业提供了可复用的技术范式,对降低大模型推理成本、推动开源生态发展具有示范意义。

技术解析

  • Kimi Delta Attention(KDA)与 Delta Rule:KDA 以固定容量白board替代传统注意力不断增长的 notebook,通过 β(写入强度)和 α(遗忘因子)两个可学习参数控制信息更新,每个注意力头配备 128 个独立遗忘调节器,实现"擦除后写入"的选择性记忆,避免新信息干扰旧记忆。
  • 混合注意力架构:K3 共 93 层注意力层,其中 69 层使用 KDA 保持内存恒定,24 层采用 Gated MLA(DeepSeek 架构的演进版)压缩 KV cache 为隐变量表示,两者以约 3:1 比例交错排列,兼顾长程高效处理与关键细节精确召回。
  • Stable LatentMoE 与稀疏激活:模型采用 896 个专家的 MoE 设计,每 token 仅激活 16 个专家,2.8 万亿总参数中仅约 1040 亿参数参与计算,通过 Quantile Balancing 方法直接从路由分数分位数设置专家偏置,实现负载均衡而无需辅助损失项。
  • 训练与优化创新:引入 Per-Head Muon 优化器使各注意力头独立适配学习率,采用自定义激活函数 SiTU 替代 GeLU/SwiGLU,并在 SFT 阶段即进行量化感知训练,最终权重以 MXFP4/MXFP8 格式发布,无需后量化即可在多种硬件上运行。
  • Attention Residuals(AttnRes):作为标准残差连接的替代方案,允许后续层直接从早期层选择性拉取表征,缓解深度网络中的 PreNorm 信息稀释问题,在增加不足 2% 计算成本的前提下提升约 25% 训练效率。

行业启示

  • 从"规模堆砌"转向"智能资源管理":K3 证明通过架构级创新(如选择性记忆、稀疏激活)可在控制计算成本的前提下突破参数规模瓶颈,为行业提供了一条可持续的大模型发展路径,减少对单纯硬件投入的依赖。
  • 开源模型进入万亿参数时代:K3 作为首个 3 万亿参数级开源模型,配合 vLLM 社区支持和自定义许可协议,将推动开源生态在长上下文、多模态 Agent 等高端场景的竞争,加速闭源与开源模型的能力收敛。
  • 长上下文实用化迎来关键突破:百万 token 上下文窗口结合高效注意力机制,使处理超长文档、代码库和复杂多步任务成为可能,将催生更多企业级 AI 应用,建议从业者关注 KDA 类机制在自身项目中的适配潜力。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Open Source 开源 LLM 大模型 Training 训练 Multimodal 多模态 Research 科学研究