A technology spin-off from the Knowledge Engineering Group (KEG) at Tsinghua University's Department of Computer Science, focused on developing foundation large models for Artificial General Intelligence (AGI), with the mission of "making machines think like humans." Listed on the Main Board of the Hong Kong Stock Exchange on January 8, 2026. The following comprehensively covers the GLM series' technical evolution, core innovations, product matrix, and major developments over the past six months (January–June 2026).
The following are technology and product updates officially released or announced in the past six months, listed chronologically. These are the current key talking points for external communications.
From the 2021 architecture foundational paper to the 2026 long-horizon task flagship model, the public release cadence and key characteristics of the GLM series are as follows.
| Model | Release Date | Key Characteristics |
|---|---|---|
| GLM Architecture Foundational Paper | March 2021 | "GLM: General Language Model Pretraining with Autoregressive Blank Infilling," arXiv:2103.10360, autoregressive blank infilling pretraining, laying the technical foundation for the entire subsequent model family |
| GLM-10B | 2021 | First 10B-parameter pre-trained large model |
| GLM-130B | August 2022 | 130B-parameter bilingual (Chinese-English) foundation model, benchmarked against GPT-3 (175B), arXiv:2210.02414 |
| ChatGLM-6B / 130B | March 2023 | China's first open-source bilingual dialogue model, cumulative downloads 30M+, GitHub Stars 150K+ |
| ChatGLM-2 | June 2023 | Context extended to 32K, introduced FlashAttention, inference speed doubled |
| ChatGLM-3 | October 2023 | Native tool calling, code interpreter, agent capabilities; on-device 1.5B/3B models |
| GLM-4 | January 16, 2024 | RMSNorm, SwiGLU, GQA architecture; 10T tokens pretraining; 60% performance improvement over previous generation, approaching GPT-4 |
| GLM-4-9B Series (4 models) | June 5, 2024 | Includes 1M context version GLM-4-9B-Chat-1M; MIT License |
| GLM-Zero / GLM-Z1 | 2024–2025 | Reasoning model series, Z1 fast version speed reaches 8x that of DeepSeek-R1, cost reduced to 1/30 |
| GLM-4-32B-0414 Series (4 models) | April 2025 | Includes GLM-Z1-Rumination-32B-0414 rumination model, benchmarked against OpenAI Deep Research; MIT License |
| GLM-4.5 / GLM-4.5-Air | July 28, 2025 | First self-developed MoE architecture (355B total / 32B active); native fusion of ARC (Agentic/Reasoning/Coding) capabilities; arXiv:2508.06471; MIT License |
| GLM-4.5V | August 11, 2025 | Visual reasoning model, SOTA on 41 multimodal benchmarks, arXiv:2507.01006 |
| GLM-4.6 | September 2025 | Context extended to 200K, token consumption reduced by ~15% compared to previous generation |
| GLM-4.7 | December 22, 2025 | 355B model, AIME 2025 reaches 95.7%, first in open-source on CodeArena |
| GLM-4.7-Flash | January 20, 2026 | 30B total / 3B active, hybrid reasoning, free to call |
| GLM-Image | January 16, 2026 | 16B, co-open-sourced with Huawei, full-flow training on domestic chips |
| GLM-5 | February 12, 2026 | 744B total / 40B active MoE; DSA sparse attention + Slime asynchronous RL framework; 28.5T tokens pretraining; arXiv:2602.15763 |
| GLM-5V-Turbo | April 2, 2026 | Native multimodal coding foundation model, arXiv:2604.26752 |
| GLM-5.1 | April 7, 2026 | Long-horizon task agent model, MIT License, officially disclosed "8-hour autonomous work" capability |
| GLM-5.1-HighSpeed | May 2026 | 400 tokens/s inference speed, global record |
| GLM-5.2 | June 13–17, 2026 | 1M (1 million token) true lossless context; MoE 744B total / ~40B active; MIT License open-source |
Core technical parameter comparison across the three flagship generations: GLM-4.5 → GLM-5 → GLM-5.2.
| Dimension | GLM-4.5 (2025-07) | GLM-5 (2026-02) | GLM-5.2 (2026-06) |
|---|---|---|---|
| Total Parameters | 355B | 744B | 744B |
| Active Parameters | 32B | 40B | 40B |
| Architecture | MoE | MoE (deeper & narrower, 80 layers / 256 experts) | MoE + Dynamic Sparsity |
| Context | 128K | 202K | 1M (1 million) |
| Pretraining Data | 23T tokens | 28.5T tokens | Training data cutoff 2025-11 |
| Attention Mechanism | Full Attention | DSA Sparse (DeepSeek Sparse Attention) | DSA + IndexShare |
| RL Training Framework | Slime Asynchronous | Slime + Fully Asynchronous Agent RL | Continued Slime Line |
| Key Technical Highlights | First native fusion of ARC capabilities | Muon Split enables MLA to match GQA-8 performance; 20B tokens match DeepSeek's 943.7B token results | IndexShare reduces per-token indexer overhead to 2.9x; improved MTP eliminates train-inference inconsistency |
| Modality | Text + Code | Text + Code | Text + Code (thinking intensity split into High/Max tiers) |
| Open-Source License | MIT | MIT | MIT |
| Official Intelligence Index | — | 50 (open-source SOTA) | 51 (3rd globally, 1st in open-source) |
Four core technologies underpinning the GLM-5 series' long-context and agent capabilities.
Uses a lightweight "flash indexer" to score historical tokens for each query, computing attention only on the top-k subset, reducing core attention complexity from O(L²) to O(Lk), enabling the model to scale to 744B parameters / 28.5T training data without sacrificing long-context understanding and reasoning depth.
Zhipu's self-developed, SGLang-native reinforcement learning post-training framework, serving as the unified foundation for RL training across the entire GLM-4.5 to GLM-5.2 series. Training (Megatron) and inference (SGLang) are decoupled as independent services, supporting asynchronous rollouts and multi-task orchestration; GLM-5.2's complete post-training took only about two days.
Every 4 sparse attention layers reuse the same indexer, reducing the per-token indexer computation for 1M context to 2.9x; combined with KVShare and end-to-end TV loss to eliminate the train-inference inconsistency of multi-token prediction (MTP) layers, this is the key to achieving "1M lossless context."
Orthogonalizes the up-projection matrix of Multi-head Latent Attention (MLA) by splitting it along attention heads, enabling MLA performance to match the traditional GQA-8 architecture, while maintaining numerical stability during pretraining without clipping attention logits.
| Model | Long Text | Code | Reasoning | Vision | Video | Image Gen | Agent | Tool Call | Open Source |
|---|---|---|---|---|---|---|---|---|---|
| GLM-5.2 (Flagship) | 1M ● | ● | ● | ○ | ○ | ○ | ● | ● | MIT |
| GLM-5.1 | 200K ● | ● | ● | ○ | ○ | ○ | ● | ● | MIT |
| GLM-5 | 202K ● | ● | ● | ○ | ○ | ○ | ● | ● | MIT |
| GLM-5V-Turbo | 200K ● | ● (Visual Coding) | ● | ● | ● | ○ | ● | ● | Partial |
| GLM-4.5V | 64K ● | ○ | ● | ● | ● | ○ | Partial | Partial | MIT |
| GLM-4.7 | 205K ● | ● | ● (AIME 95.7) | ○ | ○ | ○ | ● | ● | MIT |
| GLM-Image | ○ | ○ | ○ | ○ | ○ | ● (SOTA) | ○ | ○ | MIT |
| CogVideoX v1.5 | ○ | ○ | ○ | ○ | ● (10s 4K) | ○ | ○ | ○ | Open Source |
| AutoGLM 2.0 | ○ | ○ | Partial | ● | ○ | ○ | ● (Phone) | ● | Product |
Visual reasoning model, trained with RLCS (Reinforcement Learning with Curriculum Sampling), SOTA on 41 multimodal benchmarks, topped the Hugging Face Trending list.
16B hybrid architecture (9B autoregressive + 7B diffusion decoder), co-open-sourced with Huawei, the world's first SOTA multimodal model fully trained on domestic chips, excelling at Chinese character and knowledge-intensive image generation.
Native multimodal coding foundation model for visual programming, capable of understanding images/videos/design files/document layouts, supporting multimodal tool calls like drawing boxes, screenshots, and reading webpages.
CogView4 is the first open-source text-to-image model supporting native Chinese input; CogVideoX is an open-source text-to-video model, with v1.5 supporting stable 10-second 60fps/4K output.
End-to-end voice dialogue model, 175bps ultra-low bitrate speech tokenizer, supports real-time Chinese-English bilingual voice interaction.
Speech synthesis model, two-stage architecture (LLM generates speech tokens + Flow model converts to waveform), supports zero-shot voice cloning (3–10 second prompt).
Document parsing model, approaching top closed-source levels on the OmniDocBench benchmark, two-stage architecture (layout detection + lightweight VLM).
Speech recognition model, supports real-time accurate transcription, covering dialects and professional terminology adaptation.
From mobile GUI operations to long-horizon autonomous task execution, the public release cadence of Zhipu's agent product line.
Developer-facing MaaS platform bigmodel.cn, providing GLM series API calls, model fine-tuning, and agent development toolchain; new users receive 20M free tokens upon registration.
Developer PlatformOpen experience portal for global users, try 32B/9B series models without registration, MIT License.
International PlatformConsumer AI assistant app, supporting text/image/voice/video multimodal interaction, pre-installed on Samsung Galaxy S25 and other smart terminals in China.
Consumer ProductMulti-language AI coding assistant, supporting 100+ programming languages, integrated into mainstream IDEs, covering the full workflow of code completion, generation, and explanation.
Code AssistantDeveloper-facing coding scenario subscription plan, available in Lite/Pro/Max tiers, applicable to GLM-4.5/4.5-Air/4.6 and other series models.
Subscription ServiceSubscription service for agent scenarios, launched in March 2026, supporting AutoClaw and other products.
Subscription ServiceScientific intelligence big data mining system, the predecessor of Zhipu's technology system, developed under the leadership of Tang Jie in 2006, providing scholar relationships, paper citations, and academic trend queries.
Academic IntelligenceOfficial core technical reports sorted by publication date, all verifiable via the corresponding arXiv IDs or GitHub/Hugging Face links.
The following are pricing examples publicly released on the bigmodel.cn official website; actual prices are subject to the real-time prices on the official website.
First published 2026-07-06