GLM-4 generation (GLM-4 through GLM-4.7)

GLM (General Language Model) is the model family from Zhipu AI, a Tsinghua-adjacent Chinese frontier lab that markets its products in English under the name Z.ai. The GLM-4 generation spans the GLM-4 flagship released in January 2024 and a sequence of mixture-of-experts (MoE) successors released through 2025: GLM-4.5 and GLM-4.5-Air (July 2025), GLM-4.6 (September 2025), and GLM-4.7 (December 2025), together with multimodal variants (GLM-4V, GLM-4.5V, GLM-4.6V). Beginning with GLM-4.5, the flagship models were released with open weights under the MIT License, a shift from the partial open-weight posture of GLM-4 itself. The generation was succeeded by the GLM-5 line starting February 2026.

Lineage and history

The GLM family predates the ChatGPT era. GLM-130B (2022) was one of the first bilingual Chinese/English open-weight LLMs at the ~100B scale, contemporaneous with BLOOM and OPT. The series progressed through ChatGLM (a 2023 consumer chatbot) and ChatGLM2 and ChatGLM3, then GLM-4 (the January 16, 2024 flagship). GLM-4 introduced a 128K context window and an "All Tools" agent system that autonomously selects among a web browser, Python interpreter, and text-to-image tools; Zhipu reported it as competitive with GPT-4 on benchmarks including GSM8K (93.3) and HumanEval (78.5), and as leading on Chinese-language alignment benchmarks such as AlignBench (Source: z.ai). Zhipu is among the longer-running of the Tsinghua-adjacent Chinese frontier labs.

In 2025 the company (by then operating as Z.ai in English) shifted the flagship line to a mixture-of-experts architecture and to open weights. GLM-4.5 and GLM-4.5-Air launched on July 28, 2025; a vision model GLM-4.5V followed on August 11, 2025; GLM-4.6 on September 30, 2025; a multimodal GLM-4.6V on December 8, 2025; and GLM-4.7 on December 22, 2025, with a lightweight free-tier GLM-4.7-Flash on January 19, 2026 (Source: docs.z.ai; Source: z.ai; Source: prnewswire.com).

Training and architecture

GLM-4 is built on a dense Transformer using RMSNorm, SwiGLU, and group-query attention, trained on a multilingual corpus reported at roughly 10 trillion tokens with a Chinese/English focus, and aligned through supervised fine-tuning and RLHF (Source: z.ai).

Beginning with GLM-4.5, the flagship line adopted a mixture-of-experts architecture. GLM-4.5 has 355 billion total parameters with 32 billion active; GLM-4.5-Air has 106 billion total with 12 billion active. Both are described by Z.ai as hybrid reasoning models offering a thinking mode for complex reasoning and tool use and a non-thinking mode for instant responses — a single model with switchable inference depth, following the pattern of Qwen3's unified thinking/non-thinking modes rather than a separate reasoning model in the style of DeepSeek-R1. This single-model switchable approach has appeared across several Chinese labs; GLM-4.5's reasoning capability arrived within months of DeepSeek-R1 and OpenAI o1, one instance of the fast-follow pattern (Source: z.ai).

The GLM-4.5 technical report (arXiv 2508.06471) describes several design choices: loss-free balance routing with sigmoid gates in the MoE layers; a deeper-but-narrower configuration (fewer routed experts and smaller hidden dimension, more layers) motivated by a finding that depth aids reasoning; grouped-query attention with partial RoPE and an increased attention-head count (96 heads at a 5,120 hidden dimension) that improved reasoning benchmarks such as MMLU and BBH without lowering training loss; the Muon optimizer; QK-Norm for attention-logit stability; and a Multi-Token Prediction layer for speculative decoding. Pre-training used 15 trillion tokens of general corpus followed by 7 trillion tokens of code-and-reasoning corpus, with later domain-specific stages. Reinforcement-learning post-training, run on the company's open-sourced "slime" RL infrastructure, integrated general capabilities from GLM-4-0414 and reasoning from GLM-Z1 and added agentic coding, deep search, and tool use (Source: z.ai; GLM-4.5 technical report, arXiv 2508.06471, queued for ingest).

GLM-4.6 expanded the context window from 128K to 200K tokens and was positioned as a coding-focused flagship; GLM-4.7 was described as bringing further coding, reasoning, and agentic gains and "open-source SOTA" benchmark results (Source: z.ai; Source: docs.z.ai). GLM-series models have emphasized Chinese–English parity since GLM-130B, a training-mix difference from the English-heavy mix of most US frontier models.

Capabilities and benchmarks

Z.ai compared GLM-4.5 against models from OpenAI, Anthropic, Google DeepMind, xAI, Alibaba, Moonshot, and DeepSeek across 12 benchmarks spanning agentic, reasoning, and coding tasks, and reported GLM-4.5 ranking third overall and GLM-4.5-Air sixth in that comparison. Selected self-reported results (thinking mode where applicable):

(Source: z.ai.) On the agentic τ-bench and BFCL-v3 function-calling benchmarks Z.ai reported GLM-4.5 matching Claude 4 Sonnet, and on BrowseComp it reported 26.4% with a browsing tool, ahead of Claude 4 Opus (18.8%) and near o4-mini-high (28.3%). In an internal "CC-Bench" agentic-coding evaluation across 52 multi-turn tasks, Z.ai reported GLM-4.5 winning 53.9% against Kimi K2 and 80.8% against Qwen3-Coder, with a 90.6% tool-calling success rate, while trailing Claude 4 Sonnet (Source: z.ai).

GLM-4.6 reported clear gains over GLM-4.5 across eight public benchmarks, near parity with Claude Sonnet 4 on the extended CC-Bench (48.6% win rate) while still trailing Claude Sonnet 4.5 in coding, and roughly 15% lower token consumption than GLM-4.5 on the same tasks (Source: z.ai). One review reported GLM-4.6 scoring 82.8% on a coding benchmark against GLM-4.5's 63.3% (Source: medium.com). GLM-4.7 was reported at 73.8% on SWE-bench in independent write-ups (Source: blog.stackademic.com). As with most vendor-release benchmark claims, these figures rest largely on the developer's own reporting and were not independently replicated at release; reasoning- and coding-benchmark numbers also decay quickly as competitors update.

Availability and pricing

GLM-4's smaller GLM-4-9B variants were released openly under Zhipu's custom license, which permits research and modified commercial use with registration, while the GLM-4 flagship was offered API-only. Beginning with GLM-4.5, Z.ai released the flagship weights openly under the MIT License: GLM-4.5, GLM-4.5-Air, GLM-4.6, and GLM-4.7 weights are published on Hugging Face and ModelScope, with support for vLLM and SGLang inference (Source: z.ai; Source: github.com; Source: docs.z.ai). The models are available through the Z.ai chat interface, the Z.ai API (OpenAI-compatible), OpenRouter, and a subscription "GLM Coding Plan" that Z.ai marketed as offering Claude-level coding performance at roughly one-seventh the price with triple the usage quota, with one-click compatibility with coding agents including Claude Code, Cline, Roo Code, and Kilo Code (Source: z.ai). OpenRouter listed GLM-4.6 with a 200K context window (Source: openrouter.ai).

This open/closed history places the GLM-4 generation between the surrounding Chinese and US positions on weight release. GLM-4 itself shipped only smaller variants openly while keeping the flagship API-only; from GLM-4.5 the flagship moved to a fully permissive MIT release, aligning it with Alibaba's Qwen3 (Apache 2.0 at 235B) and DeepSeek-V3, and contrasting with the research-only license of Kimi K2 (CC BY-NC-ND) and the closed flagships of Baidu's ERNIE 4, GPT-4o, and Claude. The shift makes the later GLM-4 models a clearer data point in open-source AI and open-weight frontier model debates than GLM-4 had been, a dynamic also discussed under US–China AI competition and the US–China frontier-model comparison.

Safety and regulatory environment

Z.ai did not publish a frontier-style system card or safety case for the GLM-4 generation comparable to those of US and UK labs. For GLM-4, the company described a multi-layered approach of data cleaning, harmlessness alignment, and red-teaming, and reported a SafetyBench overall score of 87.2 (Source: z.ai). GLM models are registered under China's Generative AI Interim Measures and subject to the Deep Synthesis Provisions for multimodal outputs, with content alignment required under CAC rules. Zhipu was added to the US Entity List in January 2025, making GLM training a test case for whether export controls on compute constrain Chinese frontier development.

Reception

GLM-4.5 was covered by independent commentators including Simon Willison, who noted the Z.ai rebrand from Zhipu AI and the simultaneous release of GLM-4.5-Base, GLM-4.5, and GLM-4.5-Air (Source: simonwillison.net). Coverage of the GLM-4.5 series in machine-learning press emphasized its MIT license and broad ecosystem support as lowering the cost of access to agentic models (Source: reddit.com). GLM-4.7 reporting described Z.ai as positioning itself as "China's OpenAI" and emphasized real-world development-environment performance (Source: prnewswire.com). Independent reviewers reported using GLM-4.7 alongside Claude Opus 4.5 for coding (Source: blog.stackademic.com).

The GLM-4.5 technical report describes an open-source MoE model with 355B total and 32B activated parameters and a hybrid reasoning method supporting both thinking and direct response modes, trained on 23T tokens with expert model iteration and reinforcement learning in post-training. Reported scores: 70.1% TAU-Bench, 91.0% AIME24, 64.2% SWE-bench Verified. The efficiency claim is the emphasis — "with much fewer parameters than several competitors, GLM-4.5 ranks 3rd overall among all evaluated models and 2nd on agentic benchmarks," the agentic placement above the aggregate indicating relative strength in tool use and multi-step tasks. A 106B GLM-4.5-Air variant was released alongside.

Relationships

• developer: Zhipu AI
• instance-of: General-Purpose AI (GPAI), Open-Weight Frontier Models
• superseded by: GLM-5.2 (the GLM-5 line)
• related: Qwen3, DeepSeek-R1, DeepSeek-V3, Kimi K2, ERNIE 4 / ERNIE X1, Doubao, Open-Source AI / Open-Weight Models, Fast-Follow Problem, US-China AI Competition: Different Races, Different Metrics, China and the US Are Running Different AI Races
• depends-on: Scaling Laws, China — Interim Measures for the Management of Generative AI Services, China — Provisions on the Administration of Deep Synthesis Internet Information Services, Cyberspace Administration of China (CAC)

Sources

Primary: Z.ai blog posts for GLM-4.5 (z.ai, incl. embedded GLM-4 "All Tools" material and the arXiv 2508.06471 technical report) and GLM-4.6 (z.ai); Z.ai developer release notes (docs.z.ai); GitHub (github.com); OpenRouter (openrouter.ai). Secondary: PR Newswire (GLM-4.7), Simon Willison, Medium/Stackademic reviews, Epoch AI (epochai.substack.com) via Zhipu AI. Confidence high: core architecture, release dates, licensing, and headline benchmarks are corroborated across the official posts, release notes, and independent coverage; fast-decay benchmark figures are now mainly of historical interest as the family has been superseded by the GLM-5 line.