GLM 5.3 Flash

Z.ai's efficiency tier in the GLM 5 family: a 320B Mixture-of-Experts with 18B active, text, image, and video input, 1M context, and MIT open weights.

Start a GLM 5.3 Flash Chat GLM 5.3 Flash Benchmarks GLM 5.3 Flash Specs

1 credit

What is GLM 5.3 Flash?

GLM 5.3 Flash launched August 26, 2026, after six days on OpenRouter as stealth model Ox Alpha.

320B total, 18B active

Only 8 of 288 experts fire per token: memory of a 320B model, compute of an 18B.

Natively multimodal

Vision came from the 30T-token pre-training corpus, not a retrofitted encoder.

Flash means cheap, not fast

Artificial Analysis measures 50.2 output tokens per second, below the 67 t/s class median.

GLM 5.3 Flash specs, size, and open-source status

Architecture, weights, and licensing from the Hugging Face model card and the zai-org/GLM-5 repository.

Model card

Architecture

Total parameters: 320B
Active parameters: 18B
Context window: 1,048,576 tokens
Maximum output: 131,072 tokens
Modality: Text, image, video

Weights and licensing

Size and license

Weights license: MIT
Hugging Face: zai-org/GLM-5.3-Flash
GitHub: zai-org/GLM-5
Checkpoint size: ~328 GB (FP8)
Serving stacks: SGLang, vLLM

Model card and OpenRouter endpoints API, August 2026.

Pricing / 04

GLM 5.3 Flash pricing: list price vs launch discount

Z.ai lists $0.15 input and $0.50 output per 1M tokens. The $0.075 / $0.25 quoted everywhere is a temporary 50% launch discount.

List price

$0.15

Per 1M input

Forecast your GLM 5.3 Flash cost here, not at the promotional rate.

List price

$0.50

Per 1M output

A steep 1:3.3 ratio — cap output first.

List price

$0.03

Per 1M cached input

A fifth of fresh input — what makes long context affordable.

Independent

$0.10

Blended price

Artificial Analysis, on a 7:2:1 cache/input/output mix.

A 2x spread for identical weights: Z.AI, Novita, and GMICloud discount; others charge list.

Benchmarks / 05

GLM 5.3 Flash benchmarks: vendor-reported and independent

Z.ai publishes its own table and picks the comparisons. The GLM 5.2 delta and Artificial Analysis survive scrutiny.

Z.ai-reported

Terminal-Bench 2.1

5.3 Flash84.3

Opus 4.885.0

GPT-5.6 Terra scores 87.4; the gap widens on hard tasks.

Z.ai-reported

DeepSWE v1.1

5.3 Flash63.4

GLM 5.246.2

Same family and method, a tenth of the price. AutomationBench: 48.8 vs 26.2.

Z.ai-reported

OfficeQA Pro

5.3 Flash62.4

Document and visual reasoning, reported ahead of Opus 4.8.

Artificial Analysis

Intelligence Index

Independent, against a ~27 median for similar open-weight models.

The most reliable GLM 5.3 Flash benchmark is your own backlog.

Capabilities / 06

What GLM 5.3 Flash can do

Built for agent harnesses: reasoning control, tools, structured output, and streaming over a million-token window.

Reasoning with effort control

On by default, with an adjustable effort level.

Function calling

Tools and tool choice; your app still validates every action.

Structured output

A response-format constraint returns parseable JSON.

Streaming responses

A 1.47s time to first token beats the median; throughput does not.

Image and video input

Paste a screenshot, diagram, or screen recording with your prompt.

Million-token context

The $0.03 cached-input rate keeps later turns cheap.

Comparison / 07

GLM 5.3 Flash vs DeepSeek V4 Flash

Both are cheap tiers of a frontier family shipping MIT weights. They split on modality, reasoning, and hardware.

Z.ai

GLM 5.3 Flash

Architecture: 320B / 18B active
Input: Text, image, video
Context / output: 1M / 131K
List price: $0.15 / $0.50
Self-hosting: ~328 GB FP8

DeepSeek

DeepSeek V4 Flash

Architecture: 284B / ~13B active
Input: Text only
Context / output: 1M / 384K
List price: $0.14 / $0.28
Self-hosting: MIT weights

Both are in the glm5.app model list, so one prompt runs through each.

Release timeline / 08

GLM 5.3 Flash release timeline, from Ox Alpha to open weights

Six days as an anonymous stealth model, then a full open-weight launch.

Aug 20 — Ox Alpha appears

`stealth/ox-alpha` lands on OpenRouter with no disclosed provider, free per token.

Aug 20-26 — agent tools pile in

The five heaviest consumers are agentic coding tools; OpenCode lists it free.

Aug 26 — GLM 5.3 Flash launches

Z.ai confirms the Ox Alpha lineage; model card, MIT weights, and pricing land together.

The preview ran on Chinese chips

A custom SGLang-derived engine, reported at roughly 3x. Vendor undisclosed.

Weights shipped on Hugging Face

`zai-org/GLM-5.3-Flash` plus a BF16 variant and ModelScope mirror. No GitHub repo.

Where the news breaks

OpenRouter first on pricing, Hugging Face on quantizations. For field reports, Reddit still says Ox Alpha.

Who it is for / 09

Who uses GLM 5.3 Flash?

Teams paying flagship rates for work that never needed flagship depth — and where it is wrong.

High-volume agentic coding

Tool calling at a tenth of the GLM 5.3 token price makes long agent loops affordable.

Repository-scale work

A million-token window keeps a codebase resident, so cross-file questions skip retrieval.

Visual debugging

Broken layouts, reproduction recordings, schemas in diagrams.

Teams pricing Z.ai against rivals

Compare against DeepSeek V4 Flash and Qwen in glm5.app Chat.

Self-hosters with a cluster

MIT weights and vLLM support are real, but 328 GB of FP8 rules out one box.

Not for latency-critical UIs

At ~50 output tokens per second it trails most price peers.

How to use GLM 5.3 Flash / 10

How to use GLM 5.3 Flash

Three routes: free browser chat, the glm5.app API, or the MIT weights on your own hardware.

Chat with GLM 5.3 Flash free

Pick GLM 5.3 Flash in Chat and attach screenshots. New accounts get free credits.

Call the glm5.app API

Send OpenAI-compatible requests to `https://glm5.app/api/v1` with model ID `glm-5.3-flash`.

Or self-host the weights

Pull `zai-org/GLM-5.3-Flash` and serve it with SGLang or vLLM — ~328 GB in FP8.

Integration specimen

Start with the GLM 5.3 Flash API

An OpenAI-compatible Chat Completions endpoint. Use model ID `glm-5.3-flash` at `https://glm5.app/api/v1`.

Open API documentation

Request previewcurl

curl https://glm5.app/api/v1/chat/completions \
  -H "Authorization: Bearer $GLM5_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.3-flash",
    "messages": [{
      "role": "user",
      "content": "Read this stack trace and name the failing call site."
    }],
    "stream": true
  }'

GLM 5.3 Flash FAQ

Ox Alpha, pricing, benchmarks, parameters, Hugging Face, vLLM, DGX Spark, and DeepSeek V4 Flash.

Is GLM 5.3 Flash the same model as Ox Alpha?

Yes — it ran as `stealth/ox-alpha` for six days from August 20, 2026. Saved Ox Alpha chats are remapped automatically.

Is GLM 5.3 Flash the same as GLM 5.3?

No — different tiers, roughly a 10x price gap. Flash is the 320B-A18B multimodal model; GLM 5.3 is the text flagship.

What is GLM 5.3 Flash pricing?

Z.ai lists $0.15 per 1M input, $0.03 cached, $0.50 output. The quoted $0.075 / $0.25 is a temporary 50% launch discount.

Is GLM 5.3 Flash free?

The Ox Alpha preview was free; it is metered now. You can still try it free in glm5.app Chat.

How many parameters does GLM 5.3 Flash have, and how big is it?

320B total, 18B active per token, 45 layers routing 8 of 288 experts. The FP8 checkpoint is ~328 GB.

Is GLM 5.3 Flash on Hugging Face and GitHub?

Weights are on Hugging Face at `zai-org/GLM-5.3-Flash` under MIT. No dedicated GitHub repo — recipes live in `zai-org/GLM-5`.

Can I run GLM 5.3 Flash on vLLM or a DGX Spark?

vLLM, SGLang, TokenSpeed, and KTransformers are supported. 328 GB will not fit one DGX Spark; community reports cluster two at 4-bit.

GLM 5.3 Flash vs DeepSeek V4 Flash — which should I use?

Flash for visual input or deliberation; DeepSeek V4 Flash for a flatter 1:2 ratio or 384K output.

What does Reddit say about GLM 5.3 Flash?

Most hands-on reporting sits under the preview name, so search Reddit for Ox Alpha instead.

Start here

Start with GLM 5.3 Flash today

Run your hardest task through GLM 5.3 Flash free in the browser, then wire the model ID into your pipeline.

Start a GLM 5.3 Flash Chat Explore GLM 5.3 View Pricing

GLM 5.3 Flash

Create with GLM 5

What is GLM 5.3 Flash?

320B total, 18B active

Natively multimodal

Flash means cheap, not fast

GLM 5.3 Flash specs, size, and open-source status

Architecture

Size and license

GLM 5.3 Flash pricing: list price vs launch discount

Per 1M input

Per 1M output

Per 1M cached input

Blended price

GLM 5.3 Flash benchmarks: vendor-reported and independent

Terminal-Bench 2.1

DeepSWE v1.1

OfficeQA Pro

Intelligence Index

What GLM 5.3 Flash can do

Reasoning with effort control

Function calling

Structured output

Streaming responses

Image and video input

Million-token context

GLM 5.3 Flash vs DeepSeek V4 Flash

GLM 5.3 Flash

DeepSeek V4 Flash

GLM 5.3 Flash release timeline, from Ox Alpha to open weights

Aug 20 — Ox Alpha appears

Aug 20-26 — agent tools pile in

Aug 26 — GLM 5.3 Flash launches

The preview ran on Chinese chips

Weights shipped on Hugging Face

Where the news breaks

Who uses GLM 5.3 Flash?

High-volume agentic coding

Repository-scale work

Visual debugging

Teams pricing Z.ai against rivals

Self-hosters with a cluster

Not for latency-critical UIs

How to use GLM 5.3 Flash

Chat with GLM 5.3 Flash free

Call the glm5.app API

Or self-host the weights

Start with the GLM 5.3 Flash API

GLM 5.3 Flash FAQ

Is GLM 5.3 Flash the same model as Ox Alpha?

Is GLM 5.3 Flash the same as GLM 5.3?

What is GLM 5.3 Flash pricing?

Is GLM 5.3 Flash free?

How many parameters does GLM 5.3 Flash have, and how big is it?

Is GLM 5.3 Flash on Hugging Face and GitHub?

Can I run GLM 5.3 Flash on vLLM or a DGX Spark?

GLM 5.3 Flash vs DeepSeek V4 Flash — which should I use?

What does Reddit say about GLM 5.3 Flash?

Start with GLM 5.3 Flash today