AI News AI资讯 1h ago Updated 1h ago 更新于 1小时前 49

GLM-5.3-Flash vs Qwen3.8-Flash-Next: Two Chinese AI Labs Independently Converge on the Same Model Architecture GLM-5.3-Flash 与 Qwen3.8-Flash-Next:两家中国 AI 实验室独立趋同于相同模型架构

Z.ai and Alibaba independently converged on nearly identical architecture choices for their new frontier open-weight models: GLM-5.3-Flash (320B total / 18B active) and Qwen3.8-Flash-Next (125B total / 6B active). Both models adopt a 3:1 hybrid of linear attention to full attention layers, a compressed indexer capping context retrieval at 2048 tokens, four parallel gated residual streams, and the Muon optimizer with fused matrices split before orthogonalization. GLM-5.3-Flash achieves a 3x atten Z.ai发布GLM-5.3-Flash(320B参数,18B激活),Alibaba发布Qwen3.8-Flash-Next(125B参数,6B激活),两款模型架构高度趋同 两者均采用3:1线性注意力与全注意力混合架构,配合4倍压缩索引器(保留2048 token)实现百万级上下文高效处理 共同放弃传统单一残差流,采用四分支门控结构(GLM用mHC,Qwen用Gated Residual),并使用Muon优化器配合融合矩阵拆分策略 GLM-5.3-Flash支持原生多模态,训练30T token,价格仅为GLM-5.2的十分之一,编程与Agent性能接近Claude Opus 4.8 唯一分歧在

72
Hot 热度
68
Quality 质量
70
Impact 影响力

Analysis 深度分析

TL;DR

  • Z.ai and Alibaba independently converged on nearly identical architecture choices for their new frontier open-weight models: GLM-5.3-Flash (320B total / 18B active) and Qwen3.8-Flash-Next (125B total / 6B active).
  • Both models adopt a 3:1 hybrid of linear attention to full attention layers, a compressed indexer capping context retrieval at 2048 tokens, four parallel gated residual streams, and the Muon optimizer with fused matrices split before orthogonalization.
  • GLM-5.3-Flash achieves a 3x attention compute reduction and 4.4x KV cache savings over its full counterpart, while Qwen3.8-Flash-Next trains with roughly one-ninth the compute of Qwen3.7-Plus.
  • The sole architectural disagreement is positional encoding: GLM drops rotary position embeddings entirely in favor of implicit positional signals through recurrent linear layers, while Qwen retains RoPE.
  • Both models target 1M-token context windows and represent a significant cost-performance leap, with GLM approaching Claude Opus 4.8 on coding and agentic benchmarks at a fraction of the price.

Why It Matters

The independent convergence of two major Chinese AI labs on the same architectural recipe signals that the frontier of efficient large-language-model design is reaching a shared optimum, making these patterns highly actionable for practitioners building or fine-tuning production models. The dramatic cost reductions—whether through linear attention, sparse indexing, or the Muon optimizer—demonstrate that open-weight models can now compete with closed proprietary systems on capability while remaining economically viable at scale. For researchers, this convergence also highlights which architectural choices are robust across implementations and which remain genuinely open questions, such as positional encoding strategies.

Technical Details

  • Architecture convergence: Both models use a 3:1 ratio of linear-attention layers to full-attention layers. GLM-5.3-Flash stacks 34 Kimi Delta Attention (KDA) layers with 11 NoPE multi-head latent attention (MLA) layers; Qwen3.8-Flash-Next uses a repeating block of 3 Gated DeltaNet (GDN) layers plus 1 Qwen Sparse Attention (QSA) layer.
  • Context compression: Both compress the context 4x via a lightweight learned indexer and cap attention at 2048 tokens. GLM introduces IndexPool (weighted pooling of four indexer key vectors) to manage cost at 1M-token contexts; Qwen uses micro-block granularity scoring 4-token blocks and retaining top 512 blocks.
  • Residual stream design: Both abandon the single residual stream in favor of four parallel gated branches. GLM uses Manifold-Constrained Hyper-Connections (mHC); Qwen uses its own Gated Residual variant, which the Qwen team found ablation-equivalent to mHC while reducing memory-access overhead and enabling FP8 residual storage.
  • Optimization: Both train with the Muon optimizer, splitting fused projection matrices (QKV, SwiGLU, GDN) into independent components before orthogonalization. Qwen assigns Muon to 2-D linear maps and AdamW to embeddings, routers, and low-rank parameters, and dropped batch-size warmup after measuring it added 18.8% optimizer steps with no benefit.
  • Model specs: GLM-5.3-Flash is a 320B-parameter MoE with 18B active parameters, trained on 30T tokens, MIT-licensed, with $0.15/$0.50 per million tokens pricing. Qwen3.8-Flash-Next is a 125B main model plus 51B n-gram embedding table with 6B active parameters, native 262K context extensible to 1M via YaRN.

Industry Insight

  • The rapid independent convergence on linear attention + sparse full-attention hybrids suggests this is the emerging standard for cost-efficient frontier models; practitioners should prioritize evaluating these architectures for production deployments where 1M-token context and low inference cost are critical.
  • The Muon optimizer with fused-matrix splitting is proving to be a robust training recipe across different model scales and architectures—teams still using AdamW or standard fused-matrix Muon should consider this refinement, particularly for large-scale pretraining.
  • GLM's decision to drop rotary position embeddings entirely and rely on implicit positional encoding through recurrent layers is a bold design choice that warrants serious investigation; if subsequent models validate this approach, it could simplify architectures and reduce parameter counts across the field.

TL;DR

  • Z.ai发布GLM-5.3-Flash(320B参数,18B激活),Alibaba发布Qwen3.8-Flash-Next(125B参数,6B激活),两款模型架构高度趋同
  • 两者均采用3:1线性注意力与全注意力混合架构,配合4倍压缩索引器(保留2048 token)实现百万级上下文高效处理
  • 共同放弃传统单一残差流,采用四分支门控结构(GLM用mHC,Qwen用Gated Residual),并使用Muon优化器配合融合矩阵拆分策略
  • GLM-5.3-Flash支持原生多模态,训练30T token,价格仅为GLM-5.2的十分之一,编程与Agent性能接近Claude Opus 4.8
  • 唯一分歧在于位置编码:GLM完全移除RoPE依赖线性层隐式传递位置信息,Qwen保留RoPE

为什么值得看

两款独立开发的开源前沿模型在架构设计上惊人一致,揭示了当前大模型高效化的技术收敛路径。对从业者而言,这提供了验证过的工程范式参考,同时展示了中国AI实验室在降低训练成本与提升推理效率方面的快速迭代能力。

技术解析

  • 混合注意力架构:GLM-5.3-Flash采用34层线性注意力(Kimi Delta Attention,per-channel衰减门控)+11层全注意力(NoPE MLA);Qwen3.8-Flash-Next采用48层循环结构(3层Gated DeltaNet + 1层Qwen Sparse Attention),两者均实现3:1比例,线性层将历史压缩为固定大小状态,计算复杂度与上下文长度无关。
  • 上下文压缩机制:两者均通过轻量级索引器对上下文进行4倍压缩后评分,保留top-2048 token。GLM引入IndexPool将4个索引键向量加权池化为1个,Qwen采用micro-block粒度(4-token块)选择512个块。Qwen报告在1M token下预填充提速7.6倍、解码提速4.9倍。
  • 四分支门控残差结构:GLM采用DeepSeek起源的Manifold-Constrained Hyper-Connections(mHC),Qwen设计Gated Residual(元素级读门控+逐分支标量写门控),后者去除分支混合步骤以降低内存访问开销,并支持FP8残差存储。Qwen消融实验显示两者性能相当。
  • 优化器策略:两者均使用Muon优化器,并将融合投影矩阵(QKV、SwiGLU、GDN)拆分为独立变换后再进行正交化。Qwen将Muon分配给2D线性层,AdamW用于嵌入、路由和低秩参数,同时移除batch-size warmup(实测节省18.8%优化器步数且无性能损失)。
  • 位置编码分歧:GLM-5.3-Flash完全移除RoPE(qk_rope_head_dim=0),依赖线性层的隐式位置传递;Qwen保留RoPE。这是两者唯一明确的技术分歧点。

行业启示

  • 架构收敛信号:线性注意力+稀疏全注意力混合、多分支门控残差、Muon优化器已成为高效大模型的新标准配置,后续研究可在此基线上迭代而非重复探索。
  • 成本效率竞争白热化:GLM-5.3-Flash以十分之一价格实现接近Opus 4.8的性能,Qwen3.8训练成本仅为上代的九分之一,表明中国实验室在"性价比前沿"赛道已形成差异化优势。
  • 开源策略分化:GLM采用MIT许可直接开源,Qwen以"Next"版本预览下一代架构,两者均通过开源建立生态影响力,但GLM更侧重快速占领市场,Qwen更注重架构演进的长期布局。

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 Research 科学研究 Product Launch 产品发布