Home

AI on Demandpowered by Z.ai

GLM-5.3-Flash — a natively multimodal open-weights Mixture-of-Experts model for coding, reasoning and agentic workflows, operated by stepping stone entirely on Swiss infrastructure.

GLM-5.3-Flash is a natively multimodal open-weights model from Z.ai for text and image processing, coding, reasoning and agentic applications. Its Mixture-of-Experts architecture contains approximately 320 billion parameters, with about 18 billion active for each token. This combines high overall capacity with a compute profile designed for efficient inference.

The model combines sparse and linear attention to reduce the cost of long-context processing while preserving access to relevant information across large inputs. GLM-5.3-Flash supports a context window of up to 262'144 tokens and allows applications to control the thinking budget through different reasoning levels.

stepping stone operates GLM-5.3-Flash entirely on Swiss infrastructure. Customers access the model through an OpenAI-compatible API and can integrate it directly into existing applications, development environments and agentic systems.

GLM-5.3-Flash is suitable for organisations that need a capable model for complex technical and multimodal workloads without transferring their data to proprietary US-based AI platforms.

Typical use cases include software development, debugging, processing large information sets, image and document analysis, and agentic workflows involving external tools.

Open weights. High active efficiency. Swiss infrastructure.

GLM-5.3-Flash combines high total model capacity with only around 18 billion active parameters per token. Its hybrid attention architecture, very large context window and controllable reasoning depth make it a capable foundation for long, technical and agentic workflows.

Operation by stepping stone keeps data processing on Swiss infrastructure and provides customers with a standardised API, predictable operation and personal support from our team in Bern.

Your Consultant

Scope of services

AI model on demand

Access to GLM-5.3-Flash for coding, text and image analysis, reasoning and agentic workflows through an OpenAI-compatible API.

GPU performance on demand

Scalable GPU infrastructure in Swiss data centres for demanding inference workloads, without the need to build and operate dedicated AI infrastructure internally.

Managed service

Deployment, monitoring, maintenance and support by stepping stone. We operate the model and infrastructure so customers can focus on integrating AI into their applications and workflows.

Areas of application

Software development

GLM-5.3-Flash is designed for demanding programming and software-development tasks.

The model can analyse existing code, assist with debugging and work through multi-step development tasks. Combined with an agent harness and suitable tools, it can inspect files, propose or execute changes, run tests and incorporate the results into the next step.

Agentic workflows

GLM-5.3-Flash is suitable for workflows in which the model uses tools and pursues tasks across multiple steps.

The model can integrate external functions and systems, evaluate intermediate results and adapt its subsequent actions accordingly. Its controllable reasoning depth allows the computational effort to be matched to the complexity of each task.

Long context and multimodal analysis

GLM-5.3-Flash processes very large contexts as well as combined text and image information.

With a context window of up to 262,144 tokens, the model is suitable for large document collections, longer conversations and extensive technical contexts. Its native multimodality additionally supports joint analysis of text, images, documents and visual user interfaces.

Benchmark

The benchmarks are performed with the vLLM serving benchmark. For a defined number of input and output tokens, the test measures metrics including throughput and end-to-end latency. Because results depend on prompt length, output length, concurrency and system load, different configurations can be tested and compared. Step-by-step instructions and the required benchmark scripts can be downloaded from GitHub.

 

Call

# Set your personal key:
STONEY_KEY=sk-...

# Make key visible for vllm bench:
export OPENAI_API_KEY=$STONEY_KEY

# Start the benchmark
vllm bench serve \
 --backend openai-chat \
 --model "zai-org/GLM-5.3-Flash" \
 --base-url llm.stoney-cloud.com \
 --endpoint /v1/chat/completions \
 --dataset-name random \
 --random-input-len 1024 \
 --random-output-len 256 \
 --num-prompts 50 \
 --max-concurrency 1 \
 --percentile-metrics e2el

 

Result

============ Serving Benchmark Result ============
Successful requests:                     50        
Failed requests:                         0         
Maximum request concurrency:             1         
Benchmark duration (s):                  184.61    
Total input tokens:                      51800     
Total generated tokens:                  12800     
Request throughput (req/s):              0.27      
Output token throughput (tok/s):         69.34     
Peak output token throughput (tok/s):    142.00    
Peak concurrent requests:                2.00      
Total token throughput (tok/s):          349.93    
----------------End-to-end Latency----------------
Mean E2EL (ms):                          3691.84   
Median E2EL (ms):                        3655.69   
P99 E2EL (ms):                           4208.62   
==================================================

 

Legend

  • Successful requests: Successful prompt requests.
  • Failed requests: Unsuccessful prompt requests.
  • Maximum request concurrency: Maximum number of requests processed simultaneously.
  • Benchmark duration (s): Duration of the benchmark run in seconds.
  • Total input tokens: Total number of input tokens.
  • Total generated tokens: Total number of tokens generated by the model.
  • Request throughput (req/s): Number of requests processed per second.
  • Output token throughput (tok/s): Average number of tokens generated per second.
  • Peak output token throughput (tok/s): Highest measured number of output tokens generated per second.
  • Peak concurrent requests: Highest measured number of requests processed simultaneously.
  • Total token throughput (tok/s): Average number of input and output tokens processed per second.
  • Mean End-to-End Latency (E2EL) (ms): Average time between a request and its completed output.
  • Median E2EL (ms): Median time between a request and its completed output.
  • p99 E2EL (ms): Latency not exceeded by 99 percent of successful requests.

Price

Model
 		Context length
 		Input
CHF/MTok		Cache1
CHF/MTok		Output
CHF/MTok
GLM-5.3-Flash256k0.30000.04501.5000

1 We will able to bill the input cache tokens starting at the end of October 2026.

All prices are in CHF/MTok, excluding VAT.