"GLM 5" gets used for three different things. Sometimes it means one specific model released in February 2026. Sometimes it means the whole 5.x line, including GLM-5.1, 5.2 and 5.3. And sometimes it's simply the name of a website. If you've read a spec sheet that contradicts another one, that's usually why.
What is GLM-5? It's the open-weight foundation model Z.ai (Zhipu AI) released in February 2026. It started the GLM-5 generation. It has 744 billion total parameters with 40 billion active per token, a 200K-token context window, and MIT-licensed weights. Z.ai designed it for what it calls "agentic engineering": long, multi-step coding and agent tasks rather than quick chat. GLM-5.1, 5.2 and 5.3 are later releases built on the same generation.
This explainer is based on Z.ai's official GLM-5 model card, developer documentation and pricing page, all reviewed November 19, 2026. Benchmark figures are Z.ai's own published results. We did not re-run them. GLM5 publishes this article and operates independently of Z.ai.
GLM-5 at a Glance
| GLM-5 | |
|---|---|
| Made by | Z.ai (Zhipu AI) |
| Released | February 2026 (official repository published February 11, 2026) |
| Architecture | Mixture of experts: 744B total, 40B active parameters |
| Pre-training data | 28.5T tokens |
| Context window | 200K tokens |
| Max output | 128K tokens |
| Modalities | Text in, text out |
| License | MIT |
| API price (Z.ai) | $1.00 input / $0.20 cached input / $3.20 output per 1M tokens |
What Changed From GLM-4.x to GLM-5
GLM-5 was a generational jump rather than a point release. Its official model card gives three concrete changes compared with GLM-4.5:
- About twice the scale. It grew from 355B total parameters (32B active) to 744B (40B active), and pre-training data went from 23T to 28.5T tokens.
- Sparse attention for cheaper long context. GLM-5 integrates DeepSeek Sparse Attention (DSA). Z.ai says this greatly reduces deployment cost while keeping long-context ability.
- A new reinforcement-learning system. Z.ai built slime, an asynchronous RL infrastructure, to make post-training faster and allow more fine-grained iterations.
Here is why the "40B active" figure matters more than the 744B headline. In a mixture-of-experts model, only a slice of the network runs for each token. You still need memory for all 744B parameters, but the compute per token is closer to a 40B dense model. So GLM-5 is expensive to host but relatively efficient to run once it's loaded.
How Good Is GLM-5? Z.ai's Published Benchmarks
Against its predecessor GLM-4.7, Z.ai reports gains on nearly every benchmark it published:
| Benchmark | GLM-4.7 | GLM-5 |
|---|---|---|
| SWE-bench Verified | 73.8 | 77.8 |
| SWE-bench Multilingual | 66.7 | 73.3 |
| Terminal-Bench 2.0 (Terminus 2) | 41.0 | 56.2 (60.7 on a verified version) |
| BrowseComp | 52.0 | 62.0 |
| τ²-Bench | 87.4 | 89.7 |
| MCP-Atlas (public set) | 52.0 | 67.8 |
| Vending Bench 2 (simulated business) | $2,376.82 | $4,432.12 |
The biggest jumps are in terminal tasks, tool use through MCP and long-running agent simulations. That matches Z.ai's positioning. Z.ai's documentation goes further and says GLM-5's real-world coding usability "approaches" Claude Opus 4.5. Treat that as the vendor's claim. On Z.ai's own table, Claude Opus 4.5 still scored higher on SWE-bench Verified (80.9).
A practical reading tip: look at Terminal-Bench and MCP-Atlas if you're building agents. Look at SWE-bench if you mostly want patches to existing code. GLM-5's improvement was largest on the agent-style benchmarks.
GLM-5 vs GLM-5.1, 5.2 and 5.3
This is the most common point of confusion. In Z.ai's naming, GLM-5 is one model, and the point releases are separate models in the same generation:
| Model | Released | Context | What it added |
|---|---|---|---|
| GLM-5 | Feb 2026 | 200K | New 744B/40B-active base, DSA, slime RL |
| GLM-5.1 | Apr 2026 | 200K | Much stronger long-horizon coding; SWE-Bench Pro gains |
| GLM-5.2 | Jun 2026 | 1M | Solid 1M-token context, flexible thinking effort |
| GLM-5.3 | Aug 2026 | 1M | Post-training upgrade on GLM-5.2's base; coding and cyber gains |
If an article says "GLM 5 has a 1M-token context," it's describing GLM-5.2 or 5.3, not the original GLM-5. If a provider lists "GLM 5" without a version number, check which model ID it actually serves.
For the later releases, see our explainers on GLM-5.1, GLM 5.2 and GLM 5.3.
Is GLM-5 Open Source?
GLM-5's weights are published on Hugging Face under the MIT license, which allows commercial use, modification and redistribution as long as you keep the license notice. Z.ai also open-sourced the slime RL framework on GitHub.
What the open weights don't give you is cheap hardware. The model card lists vLLM (v0.19.0+), SGLang (v0.5.10+), KTransformers, Transformers and xLLM as supported serving stacks. A 744B-parameter checkpoint, though, needs a multi-GPU server even with quantization.
Is GLM-5 Free?
The weights are free to download. Hosted access isn't free on Z.ai's standard API: the pricing page lists GLM-5 at $1.00 per million input tokens and $3.20 per million output tokens. That's cheaper than the GLM-5.1 and 5.2 rate of $1.40/$4.40, which makes GLM-5 the budget option within the 5.x generation if 200K context is enough for you.
For the full picture across every GLM model, see Is GLM free?
Who Should Use GLM-5 Today?
GLM-5 is no longer Z.ai's flagship, but it has a clear niche:
- Use GLM-5 if you want 5.x-generation agent quality at the lowest 5.x price and your tasks fit within 200K tokens.
- Use GLM-5.2 or 5.3 if you need the 1M-token context or the strongest long-horizon coding.
- Stay on GLM-4.7 only if cost matters more than the GLM-5 gains shown above. GLM-4.7 is cheaper still at $0.60/$2.20.
Rule of thumb: start with the cheapest model that passes your five hardest real prompts, and move up only when it fails them. GLM5's API serves glm-5, glm-5.2 and glm-5.3 through one OpenAI-compatible endpoint, so comparing them is a one-word change in your code. Get the model IDs from the GLM5 API docs, or open GLM5 Chat and try a real task in your browser.
Frequently Asked Questions
What is GLM-5?
GLM-5 is Z.ai's February 2026 open-weight large language model. It's a 744B-parameter mixture-of-experts model with 40B active parameters, a 200K-token context and MIT-licensed weights, built for coding and agent tasks.
Who made GLM-5?
Z.ai, the international brand of Zhipu AI. See what GLM is for the family background.
Is GLM-5 open source?
Yes, in the open-weight sense. The weights are MIT-licensed on Hugging Face.
Is GLM-5 free?
The weights are free. Z.ai's API charges $1.00 per million input tokens and $3.20 per million output tokens.
What is GLM-5's context window?
200K tokens, with up to 128K output tokens, per Z.ai's documentation. The 1M-token context arrived with GLM-5.2.
Is GLM-5 multimodal?
No. GLM-5 is text in, text out.
The Bottom Line
GLM-5 is the model that opened Z.ai's fifth generation. It doubled the scale of GLM-4.5, added sparse attention and a new RL pipeline, and pushed open-weight agent benchmarks forward. Newer 5.x releases now outperform it, but its lower price keeps it useful for 200K-context agent work on a budget.
To see whether it's enough for your workload, try GLM-5 on a real task in GLM5 Chat, then compare it with GLM-5.2 on the same prompt.
Sources
- GLM-5 official model card — parameters, training data, DSA, slime, benchmarks and serving frameworks.
- Z.ai GLM-5 documentation — 200K context, 128K output and modalities.
- Z.ai API pricing — GLM-5 token rates.
- GLM-5 technical report (arXiv) — the official research paper behind GLM-5 and GLM-5.1.
Reviewed November 19, 2026. Benchmark scores are Z.ai's published figures, not independent tests. Prices and availability can change, so confirm them on Z.ai's official pages.



