⚛️量子位•較早收集於 56m
智譜 AI 實現每秒 400 tokens 的破紀錄推理速度
💡智譜 AI 達到每秒 400 tokens，創下頂尖 LLM 的速度紀錄，可能將改變您目前的 API 技術堆疊。
⚡ 30-Second TL;DR
有什麼變化
實現了每秒 400 tokens 的峰值推理速度
為什麼重要
此速度提升大幅降低了即時對話代理與高吞吐量程式碼助手的延遲。這為中國 AI 市場的推理效率設定了新的競爭基準。
下一步行動
將您目前對延遲敏感的應用程式與 Zhipu API 進行基準測試，評估此速度提升是否能取代更昂貴或更慢的模型供應商。
誰應關注：Developers & AI Engineers
關鍵要點
- •實現了每秒 400 tokens 的峰值推理速度
- •與現有頂尖模型相比，效能有顯著提升
- •為即時 AI 應用開發提升了效率
🧠 深度解析
Web-grounded analysis with 23 cited sources.
🔑 增強重點摘要
- •The record-breaking 400 tokens/s inference speed is specifically achieved by Zhipu AI's GLM-5.1 high-speed API, which is currently available to select enterprise clients.
- •This performance represents a 'dual breakthrough in flagship capability and ultra-low latency,' challenging the industry's previous assumption that high speed necessitates a compromise in model quality.
- •The optimization was a collaborative effort between Zhipu AI's GLM and TileRT teams, involving extensive enhancements across the inference engine, scheduling system, and underlying infrastructure.
- •Key technical improvements include rewriting the core inference pathway for increased per-card throughput, implementing dynamic batching and KV cache scheduling to reduce tail latency, and co-optimizing cluster and network performance to sustain the 400 tokens/s output.
- •The enhanced speed makes Zhipu AI's models particularly well-suited for latency-sensitive applications such as AI coding, real-time interactive agents, and real-time voice processing.
📊 競品分析▸ Show
| Model/Feature | Inference Speed (tokens/s) | Parameters (Total/Active) | Context Window (tokens) | Pricing (Input/Output per 1M tokens) |
|---|---|---|---|---|
| Zhipu AI GLM-5.1 (High-Speed API) | 400 | 754B / 40B (MoE) | 200K | $1.40 / $4.40 (GLM-5.1 API) |
| Mercury 2 | 640 (fastest output) | N/A | N/A | N/A |
| Xiaomi MiMo-V2-Flash | 150 | 309B | 256K | N/A |
| MiniMax M2 | 150 | 230B / 10B (active) | N/A | N/A |
| Zhipu AI GLM-4.7 | 55 | 400B | 200K | $0.60 / $2.20 |
| Zhipu AI GLM-4.5 | 42 | 355B / 32B (MoE) | 128K | $0.60 / $2.20 (GLM-4.5 API) |
| Cerebras (inference provider) | ~2,600 (ultra-fast) | N/A | N/A | Free tier (1M tokens/day cap) |
| GPT-5.4 (OpenAI) | N/A | N/A | N/A | $2.50 / $15 |
| Claude Opus 4.6 (Anthropic) | N/A | N/A | N/A | $15 / $75 |
| Gemini 3.1 Pro (Google) | N/A | N/A | N/A | $2.33 (Input) |
| Kimi K2.6 (Moonshot AI) | N/A | 1T | 262.1K | $0.95 / 1M tokens (cheapest in top 10) |
🛠️ 技術深入
- Zhipu AI's GLM (General Language Model) family, developed since 2019, forms the foundation of their models.
- GLM-5, a flagship model, employs a decoder-only Transformer architecture with a Mixture-of-Experts (MoE) design, featuring 744 billion total parameters and approximately 40 billion active parameters per token.
- The GLM-5 architecture incorporates DeepSeek Sparse Attention (DSA) for efficient long-context processing, RoPE positional encoding, SwiGLU activations, and post-Layer Normalization.
- GLM-4.5 also utilizes a Mixture-of-Experts (MoE) architecture, optimized for dual modes: a 'thinking' mode for complex reasoning and tool use, and a 'non-thinking' mode for faster responses.
- GLM-4.5's architecture prioritizes depth over width, featuring 96 attention heads per layer, and integrates QK-Norm, Grouped Query Attention, Multi-Token Prediction, and the Muon optimizer for improved performance.
- The 400 tokens/s speed for GLM-5.1 is powered by a high-performance inference engine, a joint development by Zhipu AI and the TileRT team.
- The TileRT engine optimizes GPU operation scheduling by restructuring it into a persistent Engine Kernel that resides on the GPU, which significantly reduces kernel startup and memory read/write delays inherent in traditional inference pipelines.
- For multi-card setups, TileRT specializes GPU nodes within an 8-card NVL topology into distinct functional Workers to enhance attention layer computation and cross-card communication efficiency.
- Zhipu AI plans further optimizations, including FP8 inference and extended context capabilities, to support even lower-latency scenarios.
- The multimodal model GLM-5V-Turbo uses a proprietary vision encoder called CogViT and employs multi-token prediction during inference to accelerate output generation.
🔮 前景展望AI analysis grounded in cited sources
Zhipu AI will likely capture a significant share of the market for real-time AI applications.
The GLM-5.1 API's record-breaking 400 tokens/s inference speed directly addresses the critical need for ultra-low latency in applications such as AI coding, real-time interaction, and voice, making their models highly competitive for these use cases.
This achievement will intensify the 'speed race' among leading LLM developers globally.
By demonstrating that high inference speed can be achieved without sacrificing model quality, Zhipu AI sets a new benchmark, compelling competitors to invest more heavily in inference optimization to maintain competitive performance and efficiency.
⏳ 時間線
2019
Zhipu AI founded as a spin-off from Tsinghua University.
2020
Began developing the GLM (General Language Model) architecture.
2022-05
Jointly debuted GLM-130B with Tsinghua University's lab.
2024-01-16
Unveiled GLM-4 model at 'Zhipu DevDay'.
2025-07
Released GLM-4.5 and GLM-4.5 Air, and rebranded internationally as Z.ai.
2026-01-08
Held IPO on the Hong Kong Stock Exchange, becoming China's first major LLM company to go public.
2026-02-11
Released GLM-5.
2026-05-22
Announced opening of GLM-5.1 high-speed API to select enterprise clients, achieving 400 tokens/s.
📎 來源 (23)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
📰
AI 週報
閱讀本週精選 AI 大事摘要 →
👉相關動態
AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源： 量子位 ↗