GLM 5.3 Flash: specs, price, benchmarks, and who sees your prompts
The Morse teamPublished 6 min read
GLM 5.3 Flash (written GLM-5.3-Flash by its maker) is an open-weight model from Z.ai, released on August 26, 2026: 320 billion parameters in total, 18 billion active per token, a context window of one million tokens, and an MIT license. Z.ai calls it the first natively multimodal model in its GLM 5 series, meaning it reads images, video, and files as well as text, and prices it at $0.15 per million input tokens and $0.50 per million output tokens on its own API.
Everything below comes from Z.ai's own pages, read on October 19, 2026: its launch post, its API guide, its price list, the model card on Hugging Face, and its privacy policy. Benchmark scores are Z.ai's, not ours.
GLM 5.3 Flash specs at a glance
| Spec | GLM 5.3 Flash | Where Z.ai says it |
|---|---|---|
| Released | August 26, 2026 | Launch post |
| Parameters | 320B total, 18B active (Mixture of Experts) | Launch post |
| Layers | 45 (GLM 4.5 had 92) | Launch post |
| Context window | 1M tokens | API guide |
| Longest reply | 128K tokens | API guide |
| Reads | Text, images, video, and files | API guide |
| Writes | Text | API guide |
| Thinking | Always on; "thinking cannot be disabled" | API guide |
| Tools | Tool calling, structured output, streaming, caching | API guide |
| API price | $0.15 in, $0.03 cached in, $0.50 out, per 1M tokens | Price list |
| License | MIT, weights on Hugging Face | Model card |
| API model name | glm-5.3-flash (and glm-5.3-flashx) | API guide |
Before launch, Z.ai ran the model anonymously as "ox-alpha" on OpenCode and OpenRouter, where it says it "quickly became the most popular model of the week," with all of that traffic "served on Chinese AI chips" (Z.ai).
How does GLM 5.3 Flash score on benchmarks?
Close to the top closed models on coding and agent work, by Z.ai's own count, and behind them on some vision tests. These are the figures in Z.ai's launch post. We haven't run them ourselves.
- DeepSWE v1.1: 63.4, against 46.2 for GLM 5.2, 58.0 for Claude Opus 4.8, and 69.6 for GPT-5.6 Terra.
- Terminal Bench 2.1: 84.3, against 85.0 for Opus 4.8 and 87.4 for GPT-5.6 Terra.
- Toolathlon Verified: 78.4, the highest in Z.ai's table, ahead of Opus 4.8 at 76.2.
- AutomationBench v1.0.6: 48.8, against 26.2 for GLM 5.2 and 52.3 for Gemini 3.7 Flash.
- Z.ai Code Bench v1.0, Z.ai's in-house test: 29.0 at max effort, against 29.5 for Opus 4.8.
- BabyVision, a visual reasoning test: 53.4, well behind Gemini 3.7 Flash at 70.9 and GPT-5.6 Terra at 61.6.
Z.ai also says the model scores 57 on the Artificial Analysis Intelligence Index v4.1.1 "at just $0.045 per task (discounted)." That figure is the cost of one benchmark task, though some summaries of the model, including the AI answer at the top of Google's results when we searched on October 19, 2026, quote it as a price per million tokens.
How much does GLM 5.3 Flash cost, and is it free?
On Z.ai's API, it isn't free. As of October 19, 2026, Z.ai's price list charges $0.15 per million input tokens, $0.03 per million cached input tokens, and $0.50 per million output tokens, with cached storage free for a limited time. Some third-party pages still quote a lower launch price, but Z.ai's own list shows no discount today. The "Flash" models that Z.ai's API serves free are older ones: GLM 4.7 Flash, GLM 4.5 Flash, and GLM 4.6V Flash.
There are two other ways to pay. FlashX, the faster route at 200 tokens a second, costs $0.37 in and $1.25 out. And Z.ai's GLM Coding Plan includes GLM 5.3 Flash with "3× the available quota compared with GLM-5.3" (Z.ai).
How is GLM 5.3 Flash different from GLM 5.3?
It's a separate, cheaper model. Z.ai says GLM 5.3 Flash "starts from a newly trained base model" and is the first GLM to mix sparse and linear attention, which cuts attention compute by 3.0 times and the memory held for long contexts (the KV cache) by 4.4 times compared with GLM 5.3 (Z.ai). On the API, GLM 5.3 costs $1.40 per million input tokens and $4.40 per million output tokens, about nine times the Flash price (price list).
Can you run GLM 5.3 Flash locally?
Yes, if you have the hardware. The weights are on Hugging Face under the MIT license, and Z.ai says it supports SGLang, vLLM, and TokenSpeed for local use, with more frameworks coming (Z.ai). The model card also links quantized builds for llama.cpp, Ollama, and LM Studio. Z.ai doesn't publish minimum hardware. By our arithmetic, 320 billion parameters at 16 bits each come to about 640 GB, so the full model is server territory even though only 18 billion parameters work on each token.
Where can you use GLM 5.3 Flash, and who sees your prompt?
The weights are the same wherever you use the model. Who runs it, and what they keep, differs a lot. Z.ai's own chat at chat.z.ai bills itself as "powered by GLM-5.3-Flash."
| Where you use it | Who can read your prompt | What's kept, per their own pages |
|---|---|---|
| Z.ai's chat, as an individual | Z.ai | Your input, "as long as you have an account" (privacy policy, last updated September 29, 2025) |
| Z.ai's API, as a business | Z.ai, while it answers | Z.ai's API addendum says it does "not store any of the content" you send, which is "not saved on our servers" (Z.ai) |
| Your own hardware | Only you | Whatever your machine keeps |
| Morse, in E2EE mode | Only a secure enclave your browser has verified | Morse's servers relay text they can't read; your history stays encrypted in your browser |
Z.ai's consumer policy says your data is "generally processed in Singapore," that it may go to affiliates in the Z.ai Group, and that Z.ai may share it with "any competent law enforcement body, regulatory, government agency, court or other third party" where it believes disclosure is necessary under the law, to defend legal rights, or to protect someone's vital interests. Deleting a conversation removes it from your history at once and from Z.ai's back end automatically, the policy says, without giving a time frame. Z.ai deserves credit for putting a no-storage promise for API content in writing.
How to use GLM 5.3 Flash privately on Morse
In Z.ai's chat, Z.ai's servers read your prompt and keep it with your account. On Morse, an E2EE chat can be read only by a secure enclave your browser has verified, and in every mode your chat history lives only on your device. GLM 5.3 Flash is one of the models on Morse's Free plan, and on Pro, at $19 a month, it also runs in end-to-end encrypted chats.
- E2EE, for prompts you wouldn't hand to any host. Start a chat in E2EE and pick GLM 5.3 Flash. Before anything leaves your device, your browser checks the secure enclave's hardware attestation against versions pinned in the app, then encrypts the message so only that enclave can open it. Morse's servers pass along ciphertext they can't read, and the reply is decrypted only in your browser. Each reply carries a lock you can open to see every check that passed. Third-party providers run the enclaves, not Morse, and the guarantee rests on the enclave hardware and attestation working as designed. E2EE chats use only tools that run on your device, so for web search, switch to another mode.
- Private, for everyday chats with open models and web search. Open-weight models like GLM, Kimi, and DeepSeek run on hosts that agreed to keep nothing. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't; otherwise Morse relies on the provider's commitment.
- Anonymous, for GPT and Claude. Proprietary models go through Morse's proxy, which removes your account, email, and IP address, so the provider gets the question without knowing it's you. The provider may keep it for a limited time to check for abuse, under its own policy.
Whatever the mode, your chat history is stored only in your browser, encrypted, and Morse doesn't train AI on your chats, sell your data, or show ads (privacy policy). The Free plan costs nothing and needs no card: 15 messages a day shared across GPT-6 Luna, Claude Haiku 5.5, GLM 5.3 Flash, and more. For how the enclave check works step by step, read how end-to-end encrypted AI chat works, and for the wider picture, our guide to using AI privately.
Start free and chat with GLM 5.3 Flash today, or open an E2EE chat with it on Pro.
Common questions
Can you turn off thinking in GLM 5.3 Flash?
No. Z.ai's API guide says the model's thinking setting "only supports enabled" and that "thinking cannot be disabled." You can lower the effort instead: the Hugging Face model card lists low, high, and max for reasoning_effort, with max as the default.
Does GLM 5.3 Flash support function calling?
Yes. Z.ai's guide lists tool calling, structured JSON output, streaming, and context caching among its capabilities, and recommends turning on tool_stream along with stream for streaming requests.
What is GLM 5.3 FlashX?
It's a faster way to call the same model on Z.ai's API, which Z.ai says runs at 200 tokens a second. As of October 19, 2026 it costs $0.37 per million input tokens and $1.25 per million output tokens, and Z.ai says it isn't yet part of the GLM Coding Plan.
Who makes GLM 5.3 Flash?
Z.ai, the company behind the GLM models. The privacy policy for its consumer services names JINGSHENG HENGXING TECHNOLOGY PTE.LTD, registered in Singapore, as the company responsible for your data.