GLM 5.3 Flash
Z.ai
- Parameters
- 320B
- Context
- 1M tokens
- Input
- Text, Images, Video
- License
- MIT
glm-5.3-flashModel card for GLM 5.3 Flash ↗Call them per token on Inco Platform, run them on your own compute with Inco Splash,
or on capacity dedicated to your team with Inco Endpoints.
Zero data retention is available whichever product you choose. On Platform and Endpoints, you decide how your data is handled, all the way to prompts and outputs not being kept once the request is served. With Splash, they never leave your machine.
Serverless open-weight models, at the speed your workload needs
The leading open-weight models, per token, through the OpenAI- and Anthropic-compatible APIs you already use. Run each model at the tier that fits the job.
An illustration of the Inco Platform console for GLM 5.3 Flash, with a tier switch and the code to call it.
from openai import OpenAIclient = OpenAI( base_url="https://api.inco.ai/v1", api_key="INCO_API_KEY",)resp = client.chat.completions.create( model="glm-5.3-flash:fast", messages=[{"role": "user", "content": task}],)
| Speed | |||
|---|---|---|---|
| Price | Half the list price | List price per token | 2× the list price |
| How to call | model-id | model-id | model-id |
| Best for | Evaluations, dataset processing, and other offline jobs | Most applications, where cost matters more than peak speed | Agents and real-time products, where each step waits on the last answer |
Each model's prices are on the live catalog. For your own speed and price, see custom models on Inco Endpoints.
Local AI, built around the model
Splash is our open-source inference engine for local compute. Its kernels, draft model, and memory plan are specialized for each model it serves. It runs on Apple silicon today, with more hardware coming soon.
Qwen3.6-35B-A3B on a 48 GB M5 Pro, with coding prompts from NVIDIA's SPEED-Bench, as published in the Splash post
brew install incoai/tap/splash splash serve --model unsloth/Qwen3.8-27B-GGUF:UD-Q4_K_M
Splash needs an M3 or newer Mac on macOS 26.4 or later, and Homebrew. This model needs at least 36 GB of unified memory.
Dedicated throughput for your team's workload
An illustration of a dedicated endpoint's dashboard: a committed token throughput reserved for your team, with your traffic running under it, above 1B tokens per minute.