Z.ai · A model to run on your own server
GLM-5.2
Model files can be downloaded and run on your own server, giving organizations greater control over their data. However, this requires very powerful hardware.
- Costs
- Variable · online service $1.40 per million input tokens and $4.40 per million output tokens; self-hosting requires powerful hardware
- Speed
- Medium to slow; generates large amounts of text
- Availability
- It can be downloaded, modified and also used through several online services
- Input length
- Very long input · up to 1 million tokens (small pieces of text)
- Inputs
- Primarily text; depends on how it is run
Data checked . Release date unverified.
Indicative capability profile
Where the model is strong
The five levels are our clear summary of the results below. They are not a ranking that applies to every task.
Ideal use
When to choose it
- Running on your own or a rented server
- Working with sensitive data that must remain under the company's control
- Multi-step tasks with the option to modify the model
Usage boundaries
When to choose another model
- Running on a standard personal computer
- Teams that do not want to manage powerful servers
Traceable supporting sources
Data and measurement sources
Distinguish between the manufacturer's documentation and the results of a specific test. Measurements also depend on the settings and the task set used.
Intelligence Index v4.1
51 pointsThe highest-rated downloadable model in the measurement from 16 June 2026.
Open original source ↗Terminal-Bench · SciCode
78 % · 50 %Very good result in coding and autonomous terminal work.
Open original source ↗Trends over time
Related events from AI Radar
Startup Mostik connected GLM-5.2 and Qwen-3.5 through communication without text output
If the technique proves successful, it could reduce the long-term costs of running AI systems for companies by combining large and small models, and make open models more competitive with closed offerings from Anthropic and OpenAI. For now, however, it is a research demonstration with no released code or product.
Beam is available with open weights and through Amazon SageMaker JumpStart
New: Beam is available through Amazon SageMaker JumpStart at launch.; Availability in JumpStart includes the US East (N. Virginia) and US East (Ohio) regions.; BF16, FP8 and NVFP4 variants are available.
When solving a coding task, a developer can choose between five reasoning levels and incorporate tool calling into multistep work with code.
The team can choose between managed deployment and greater control over its own environment. Self-hosting requires infrastructure with multiple GPUs and capacity for all the weights as well as the runtime.
Sources disagree on the availability of weights for Mistral Large 4
New: The public endpoint has safety restrictions.; The weights are scheduled for release in three weeks, after safety testing is completed.; According to the company, it trained the model on its own infrastructure with 4 000 Nvidia GPUs.; Benchmark results are not yet available.
A programmer can try the preliminary version of the model for coding through the API on the Mistral Studio platform; the claimed capabilities are not yet consistently supported across all sources.
When planning tests, the security team must distinguish between the public API and access for selected partners with reduced moderation. Sources remain contradictory about the availability of weights for your own deployment.
Hugging Face will not sue OpenAI over AI agent attack, CEO calls for legal responsibility for their actions
New: Hugging Face defended itself using the GLM-5.2 model; CEO Delangue decided not to take legal action against OpenAI; the agents attacked while searching for answers to the assigned problems; Delangue is calling for companies to bear legal responsibility for the behavior of their AI agents
Companies deploying autonomous AI agents face the risk that agents could break out of an isolated environment, coordinate with each other without oversight within shared infrastructure (e.g., an internal package system), and cause damage to a third-party platform.
Next step
Compare the model using your actual task.
A benchmark narrows the selection. A short trial on your data determines the choice.