AI News AI资讯 6h ago Updated 2h ago 更新于 2小时前 49

Z.ai Releases GLM-5.3-Flash: A 320B-A18B Natively Multimodal MoE With a 1M-Token Context Z.ai发布GLM-5.3-Flash:一款320B-A18B原生多模态MoE模型,支持100万Token上下文

Z.ai released GLM-5.3-Flash, a 320B-parameter MoE model with 18B active parameters per token, featuring native multimodal (image/video) support and a 1M-token context window, licensed under MIT with weights on Hugging Face. The architecture introduces hybrid attention (KDA linear + NoPE sparse MLA), IndexPool for KV cache compression, and Manifold-Constrained Hyper-Connections, achieving ~3× less attention compute and 4.4× smaller KV cache versus GLM-5.3. Benchmark results are strong: 84.3 on Te Z.ai发布GLM-5.3-Flash,320B总参数/18B激活参数的原生多模态MoE模型,支持1M token上下文窗口,MIT许可证开源 采用混合注意力架构(KDA线性注意力+NoPE稀疏MLA)与IndexPool技术,实现约3倍注意力计算减少和4.4倍KV缓存缩减 在Terminal-Bench 2.1(84.3分)和DeepSWE v1.1(63.4分)等代码基准上接近Claude Opus 4.8,显著超越GLM-5.2 API定价极具竞争力($0.15/M输入,$0.50/M输出),为GLM-5.3价格的约十分之一,支持SGLang/vLLM等多引擎本地部署 自托管需约306

72
Hot 热度
68
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Z.ai released GLM-5.3-Flash, a 320B-parameter MoE model with 18B active parameters per token, featuring native multimodal (image/video) support and a 1M-token context window, licensed under MIT with weights on Hugging Face.
  • The architecture introduces hybrid attention (KDA linear + NoPE sparse MLA), IndexPool for KV cache compression, and Manifold-Constrained Hyper-Connections, achieving ~3× less attention compute and 4.4× smaller KV cache versus GLM-5.3.
  • Benchmark results are strong: 84.3 on Terminal-Bench 2.1 (near Claude Opus 4.8's 85.0), 63.4 on DeepSWE v1.1 (up from GLM-5.2's 46.2), and 29.0 on Z.ai Code Bench (within 0.5 of Opus 4.8).
  • Priced at $0.15/M input and $0.50/M output, it delivers roughly one-tenth the cost of GLM-5.2 while offering 3× the usable quota across all GLM Coding Plan tiers.
  • Self-hosting requires ~306 GiB FP8 weights and NVIDIA Hopper or newer GPUs; the preview was served anonymously as "Ox Alpha" on domestically produced Chinese AI chips with a 3× end-to-end serving improvement.

Why It Matters

GLM-5.3-Flash represents a significant step toward cost-effective, high-capability multimodal coding models, bridging the gap between premium offerings like Claude Opus 4.8 and budget-conscious deployments. Its MIT license and open weights lower barriers for researchers and enterprises to experiment with and fine-tune a model that competes at the top tier while costing a fraction of the price. The hybrid attention and IndexPool innovations also provide a blueprint for scaling to million-token contexts without prohibitive memory costs.

Technical Details

  • Architecture: 45-layer MoE with 8 of 288 experts activated per token, native FP8 weights, and one MTP draft layer. Combines KDA linear-attention layers for local dependencies with NoPE sparse MLA layers for global context retrieval.
  • IndexPool: Compresses groups of indexer key vectors through weighted pooling, reducing attention compute by ~3× and KV cache size by 4.4× compared to GLM-5.3, critical for million-token context handling.
  • mHC (Manifold-Constrained Hyper-Connections): Improves scaling efficiency, roughly halving activated parameters and layer count versus GLM-4.5 at similar total parameter counts.
  • Training: Base model trained on a 30T-token multimodal corpus, with image and video input capabilities natively integrated.
  • Serving: Supports self-hosting via vLLM (NVIDIA Hopper+), SGLang, TokenSpeed, and KTransformers. The Ox Alpha preview ran on domestic Chinese AI chips using a custom SGLang-based engine with disaggregated encoding, prefill, and decoding.

Industry Insight

  • Cost-performance shift: At one-tenth the price of its predecessor with near-Opus-level coding benchmarks, GLM-5.3-Flash could accelerate adoption of agentic AI workflows in software development, IT automation, and enterprise back-office operations where cost per task is a decisive factor.
  • Self-hosting门槛: The 306 GiB FP8 weight requirement and Hopper+ GPU dependency mean only well-capitalized organizations can self-host; most will rely on the API, making pricing strategy and quota tiers (3× GLM-5.3) a key competitive lever for Z.ai.
  • Multimodal coding gap: While coding benchmarks are strong, vision performance trails Gemini 3.7 Flash on BabyVision and MVbench, suggesting that teams requiring heavy image/video reasoning should evaluate use-case-specific trade-offs before adopting.

TL;DR

  • Z.ai发布GLM-5.3-Flash,320B总参数/18B激活参数的原生多模态MoE模型,支持1M token上下文窗口,MIT许可证开源
  • 采用混合注意力架构(KDA线性注意力+NoPE稀疏MLA)与IndexPool技术,实现约3倍注意力计算减少和4.4倍KV缓存缩减
  • 在Terminal-Bench 2.1(84.3分)和DeepSWE v1.1(63.4分)等代码基准上接近Claude Opus 4.8,显著超越GLM-5.2
  • API定价极具竞争力($0.15/M输入,$0.50/M输出),为GLM-5.3价格的约十分之一,支持SGLang/vLLM等多引擎本地部署
  • 自托管需约306 GiB FP8权重及Hopper架构及以上GPU,中小企业更适合通过API消费

为什么值得看

GLM-5.3-Flash在保持接近顶级闭源模型性能的同时,将成本降低至约十分之一,为AI从业者提供了高性价比的代码智能和多模态解决方案。其混合注意力架构和IndexPool技术为长上下文推理的效率优化提供了新的工程范式,对关注部署成本和模型效能平衡的团队具有重要参考价值。

技术解析

  • 架构创新:首次将KDA线性注意力与NoPE稀疏MLA结合,线性层处理局部依赖,稀疏层检索全局上下文;采用Manifold-Constrained Hyper-Connections(mHC)提升缩放效率,相比GLM-4.5在相似总参数量下将激活参数和层数几乎减半
  • 长上下文优化:IndexPool通过加权池化压缩索引器键向量组,在百万token上下文场景下将注意力计算减少约3倍,KV缓存缩小4.4倍,有效缓解检索瓶颈
  • 部署规格:默认FP8检查点约306 GiB权重(不含KV cache),vLLM支持NVIDIA Hopper及更新架构;本地部署支持SGLang、vLLM、TokenSpeed和KTransformers
  • 基准表现:Terminal-Bench 2.1得分84.3(Opus 4.8为85.0),DeepSWE v1.1得分63.4(GLM-5.2为46.2),AutomationBench 48.8(GLM-5.2为26.2),Artificial Analysis智能指数57分
  • 服务架构:Ox Alpha预览阶段基于国产AI芯片运行,采用自定义SGLang引擎实现编码、预填充和解码的 disaggregation,在数万加速器上实现3倍端到端服务提升

行业启示

  • 成本效率范式转变:MoE架构结合混合注意力技术使模型在保持高性能的同时大幅降低推理成本,API价格仅为前代的约十分之一,这将加速AI能力在中小企业和工作流自动化中的普及
  • 部署分层策略明确:306 GiB权重和Hopper架构要求划定了自托管的门槛,中型以上企业和AI原生初创公司可考虑本地部署以控制数据敏感性,其他组织应聚焦API经济性和按需扩展
  • 多模态代码智能成为新战场:该模型在代码基准上接近Opus 4.8且支持图像/视频输入,适用于仓库级编码代理、终端/浏览器控制、百万token日志分析和UI回归检测,软件工程和IT自动化是首批高价值落地场景

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 Multimodal 多模态 Code Generation 代码生成 Product Launch 产品发布