Z.ai
GLM-5.3-FlashX
glm-5.3-flashx
Total Context
1M
Max Output
131.1K
Released
N/A
Input$0.2857 / M tokens
Output$1 / M tokens
GLM-4.5
glm-4.5GLM-4.5 is an agent-oriented MoE model from Z.ai, described with 355B total parameters and 32B activated parameters. Official docs highlight reasoning, coding, tool use, and browser-style agent abilities, with both thinking and non-thinking modes. It works well as the GLM line’s earlier agent foundation model before GLM-5.
Context Window
131.1K tokens
Maximum Output
98.3K tokens
Release Date
Jul 28, 2025
Modalities
GLM-4.5 Pricing
| Input Price | Output Price |
|---|---|
| $0.4286/M | $2/M |
GLM-4.5 API Capabilities
Reasoning
Supported
Tool calling
Supported
Temperature parameter
Supported
Attachments
Not supported
Knowledge Base
2025-04-01
Endpoint Protocols
Completions APIMessages APIgemini
GLM-4.5 Model Highlights
GLM-4.5 is an open-weight text model that unifies reasoning, coding and agent capabilities with hybrid thinking modes and a 128K context.
Unified Agentic Reasoning
Combines reasoning, coding and agent-oriented training in one model for multi-step problem solving and software tasks.
Hybrid Thinking Modes
Offers a thinking mode for complex reasoning and an immediate-response mode for simpler workloads, allowing developers to match effort to the task.
Open-Weight MoE
Uses a mixture-of-experts architecture with 355B total and 32B active parameters, released under the MIT license for independent deployment.
GLM-4.5 Use Cases
GLM-4.5 suits software engineering agents, complex text reasoning and self-hosted development workflows that benefit from selectable thinking depth.
Software Engineering Agent
Plan code changes, inspect project files and carry out implementation and debugging steps across a multi-turn development task.
Complex Text Reasoning
Analyze long technical requirements or multi-step logic problems in thinking mode and produce a reasoned textual solution.
Private Model Deployment
Deploy the MIT-licensed open weights in controlled infrastructure for internal coding, reasoning or agent workloads.
How to Use GLM-4.5 via the TokenHub API
import OpenAI from "openai"
const client = new OpenAI({
apiKey: process.env.TOKENHUB_API_KEY,
baseURL: "https://us-api.tokenhub.com/v1",
})
const result = await client.chat.completions.create({})
console.log(result.choices[0]?.message?.content)GLM-4.5 Benchmarks
GLM-4.5 (Reasoning)
| Index score | ||
|---|---|---|
| Artificial Analysis Intelligence Index | Artificial Analysis broad capability aggregate | 19.5 |
| Artificial Analysis Coding Index | Artificial Analysis software task aggregate | 26.3 |
| Artificial Analysis Math Index | Artificial Analysis math reasoning aggregate | 73.7 |
| Knowledge & Reasoning | ||
| MMLU-Pro | Advanced multi-task knowledge | 83.5% |
| GPQA | Advanced science problem solving | 78.2% |
| HLE | Broad expert-level exam set | 12.2% |
| Coding & Engineering | ||
| LiveCodeBench | Live coding problems | 73.8% |
| SciCode | Scientific coding challenges | 34.8% |
| Terminal-Bench Hard | Hard terminal task execution | 22.0% |
| Math | ||
| MATH-500 | Advanced math problem solving | 97.9% |
| AIME | Competition math problems | 87.3% |
| AIME 2025 | Competition math problems | 73.7% |
| Instruction Following & Agent Tasks | ||
| IFBench | Prompt constraint adherence | 44.1% |
| AA-LCR | Long-context reasoning | 48.3% |
| τ²-Bench | Agent workflow tasks | 43.0% |
Metrics sourced from Artificial Analysis
GLM-4.5 FAQ
GLM-4.5: capabilities, use cases, limits, and TokenHub guidance.
What is GLM-4.5?+
GLM-4.5 is a Z.AI model for reasoning, coding, and native agent workflows.
Which workloads suit GLM-4.5?+
Best for code reasoning, agent workflows and tool-heavy automation, especially when deep reasoning is the priority.
Which feature stands out?+
Key strength: a unified focus on reasoning, coding, and native agents.
When should teams avoid GLM-4.5?+
It belongs to an older generation and may lack newer capabilities. For the latest capabilities matter, consider GLM-5.
What should I verify in TokenHub?+
Confirm TokenHub availability; prefer the current successor for new work.
Ready to use GLM-4.5?
Use one API key to access GLM-4.5 and more AI models through TokenHub.
Media and Discussions
Selected public videos and posts related to this model.
X (Twitter)
Z.ai ·
Introducing GLM-4.5 and GLM-4.5 Air: new flagship models designed to unify frontier reasoning, coding, and agentic capabilities. GLM-4.5: 355B total / 32B active parameters GLM-4.5-Air: 106B total / 12B active parameters API Pricing (per 1M tokens): GLM-4.5: $0.6 Input / $2.2 https://t.co/ummGTd6QLl
https://pbs.twimg.com/media/Gw8pU-_aQAA6e-8.jpgView originalZ.ai ·
To better demonstrate the coding capabilities of GLM-4.5, we have developed a coding agent inspired by Claude Code. By providing a basic full-stack website boilerplate, the agent enables users to create an entire website with just a few words. /1 Build an interactive Pokémon https://t.co/Lcih0YCOO7
https://pbs.twimg.com/amplify_video_thumb/1949822920483803137/img/OQeApuV3oA8ABbSv.jpgView originalZ.ai ·
GLM-4.5 on https://t.co/gocggreDem now features comprehensive vision capabilities. For everyday tasks, simply upload images or videos to GLM-4.5 for instant analysis and discussion. For advanced visual reasoning, switch to the specialized "GLM-4.5V" model to unlock enhanced https://t.co/zk0apKBMiP
https://pbs.twimg.com/amplify_video_thumb/1955076631438970880/img/hTvS6YaX2QyG5VpD.jpgView originalReddit
What’s your experience with GLM-4.5? Pros and cons?
Middle-Copy4577 ·
I’ve been using it alongside Claude Code, and in my experience it handles most ordinary coding tasks flawlessly. I’m curious how it stacks up against other models in terms of reasoning depth, code quality, and ability to handle edge cases.
View originalGLM 4.5 Collection Now Live!
Lowkey_LokiSN ·
https://huggingface.co/collections/zai-org/glm-45-687c621d34bda8c9e4bf503b
https://huggingface.co/collections/zai-org/glm-45-687c621d34bda8c9e4bf503bView originalDaily driving GLM 4.5 for 10 days, kinda insane how good it is at half the size of other frontier models
susmitds ·
Read full post
I’ve been running GLM 4.5 (355B) locally for about 10 days, and it’s basically replaced my old setup. I used to juggle GPT-4o/4.1 for general tasks and o3 for heavier reasoning, but after GPT-5 struggled with long research paper convos I moved to GLM 4.5, and it covers both use cases in one. Using Unsloth’s GGUF builds in llama.cpp: ud-iq2_m for 128k context (K/V cache q8) and ud-q2_k_xl when 30k okay. What’s impressive is how well it holds up under 2-bit quant + cache quant. I expected obvious degradation, but it’s solid. Non-thinking mode feels like GPT-4o in fluency, and thinking mode sits between o4-mini-high and o3—better than most open weights I’ve tried at this scale. Benchmarks aside, real usage matches the claims. I’ve thrown workloads at it I’d normally reserve for GPT-4.1 or o3, and it keeps up or beats them, without the usual quantization artifacts. Knowledge depth and overall competence are ahead of GPT-4.1/4o in many areas, and it’s one of the most stock, uncensored releases I’ve seen. Running on RTX 6000 Ada + 6000 Pro Blackwell with a llama.cpp fork for GLM 4.5 tool support. Frontend is Open WebUI with some filter functions to toggle thinking on/off, plus a fixed chat template so /nothink only applies to the last message (avoids re-prefill issues). First time I’ve daily-driven a local model this large and felt it genuinely outclass closed ones.
YouTube
GLM 4.5: This AI Model From China Is Changing Everything!
Better Stack
View originalZ.ai GLM 4.5 - Explained simply
Kashi DHQ
View originalIs the new GLM-4.5 agent-focused model overhyped?
What's AI by Louis-François Bouchard
View original