Z.ai has launched GLM-5.3-Flash, a new 320-billion-parameter multimodal artificial intelligence model designed to deliver frontier-level capabilities at a fraction of the cost of larger flagship systems. The model has 320 billion total parameters but activates only 18 billion parameters per token, and Z.ai says it can outperform its previous GLM-5.2 model across benchmarks and real-world workloads while costing roughly one-tenth as much.
The release also solves the mystery surrounding Ox Alpha, an anonymous AI model that became highly popular on OpenRouter and OpenCode before Z.ai revealed that it was a preview version of GLM-5.3-Flash. The new model is the first natively multimodal model in the GLM-5 series, supporting text, images and video, while offering a context window of up to 1,048,576 tokens. Z.ai has released the model weights publicly and is positioning GLM-5.3-Flash as a lower-cost alternative for coding, AI agents and long-running workloads.
GLM-5.3-Flash At A Glance
The new model combines a large overall parameter count with a much smaller active parameter count, allowing Z.ai to target lower inference costs without positioning the model as a lightweight system.
Key Details
| Particular | Details |
|---|---|
| Model | GLM-5.3-Flash |
| Developer | Z.ai |
| Total parameters | 320 billion |
| Active parameters | 18 billion |
| Architecture | Mixture of Experts |
| Multimodal | Text, image and video |
| Context window | 1,048,576 tokens |
| Model series | GLM-5 |
| License | MIT |
| API input price | $0.15 / 1M tokens |
| API output price | $0.50 / 1M tokens |
| Cached input | $0.03 / 1M tokens |
| Main focus | Coding, agents and long-context workloads |
| Anonymous preview | Ox Alpha |
Z.ai says the model was developed from a newly trained base model, with its architecture and training process redesigned around capability and efficiency.
Z.ai Reveals The Identity Of Ox Alpha
Before the official release, developers had been testing an anonymous model called Ox Alpha on OpenRouter and OpenCode.
The model quickly attracted attention because of its coding and agentic performance.
Z.ai later confirmed that Ox Alpha was actually GLM-5.3-Flash, which the company had released anonymously to collect unbiased feedback from developers before the public launch.
GLM-5.3-Flash
↓
Anonymous Preview
↓
“Ox Alpha”
↓
OpenRouter + OpenCode
↓
Developer Feedback
↓
Official Z.ai ReleaseZ.ai said the anonymous model became the most popular model of the week on the platforms during testing, with the traffic served on Chinese AI chips.
320 Billion Parameters, But Only 18 Billion Active
The headline specification of GLM-5.3-Flash is its 320B-A18B architecture.
The model contains 320 billion parameters in total, but only approximately 18 billion are activated for each token.
This is a mixture-of-experts approach that allows the model to maintain a large pool of learned parameters without performing the full computational workload on every token.
Parameter Structure
| Metric | GLM-5.3-Flash |
|---|---|
| Total parameters | 320B |
| Active parameters | 18B |
| Active share | 5.6% |
| Context window | 1M tokens |
| Architecture | MoE |
The distinction is important because total parameter count and inference computation are not the same thing.
320B Total Parameters
↓
Routing System
↓
18B Active Parameters / Token
↓
Lower Computation
↓
Lower Inference CostGLM-5.3-Flash Is Natively Multimodal
GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series.
It can process:
- Text
- Images
- Video
Z.ai says the multimodal capabilities are designed particularly for coding and agent applications where an AI system needs to understand what is displayed on a screen rather than relying only on text descriptions.
Multimodal Capabilities
| Input | Potential Application |
|---|---|
| Text | General AI tasks |
| Images | Visual reasoning |
| Video | Long-video understanding |
| Screenshots | Computer-use agents |
| Charts | Data analysis |
| Web pages | Browser agents |
| UI designs | Code generation |
This allows the model to connect visual information with code and tool-based actions.
One Million Token Context Window
Another major feature is the model’s 1,048,576-token context window.
A million-token context can allow an AI system to process extremely large amounts of information in a single session.
Potential applications include:
- Large software repositories
- Long videos
- Extensive documentation
- Multi-file coding projects
- Enterprise records
- Long-running AI agents
Large Context
↓
More Code + Documents + History
↓
Longer Agent Sessions
↓
Less Need To Summarize Or Reset ContextZ.ai says the architecture was specifically redesigned to reduce the cost of serving long contexts.
Hybrid Attention Cuts Long-Context Costs
GLM-5.3-Flash introduces a hybrid architecture combining sparse attention and linear attention.
Z.ai says the architecture is designed to reduce the computational burden associated with very long contexts while retaining the ability to retrieve important information.
The company also introduced IndexPool, which reduces the cache required by its sparse-attention indexer.
Attention Architecture
| Technology | Role |
|---|---|
| Linear attention | Efficiently processes local information |
| Sparse attention | Retrieves important global context |
| Lightweight indexer | Identifies relevant context |
| IndexPool | Reduces indexer cache requirements |
| Hybrid architecture | Balances efficiency and retrieval |
According to Z.ai, compared with GLM-5.3, the new architecture reduces attention computation by approximately 3.01 times and KV-cache size by approximately 4.44 times.
Why Long-Context Efficiency Matters
Long-context AI can become expensive because the model has to process increasingly large amounts of information as a session grows.
This is particularly relevant for coding agents.
An AI coding agent may repeatedly read:
- Source code
- Previous conversations
- Terminal outputs
- Documentation
- Tool results
- Test results
The context can therefore grow continuously during a long task.
Coding Task
↓
Read Repository
↓
Write Code
↓
Run Tests
↓
Read Errors
↓
Modify Code
↓
Run Tests Again
↓
Growing ContextReducing attention and memory requirements can therefore have a direct impact on the economics of agentic AI.
Z.ai Targets Claude Opus-Level Coding Performance
Z.ai says GLM-5.3-Flash approaches Claude Opus 4.8 on several coding and agentic evaluations.
On Z.ai’s internal Z.ai Code Bench v1.0, the model scored 29.0 at maximum effort, compared with 29.5 for Claude Opus 4.8.
Selected Coding Benchmarks
| Benchmark | GLM-5.3-Flash | Claude Opus 4.8 |
|---|---|---|
| Terminal-Bench 2.1 | 84.3 | 85.0 |
| Z.ai Code Bench | 29.0 | 29.5 |
| DeepSWE v1.1 | 63.4 | 58.0 |
| AutomationBench | 48.8 | 41.0 |
These are primarily Z.ai’s published comparisons and should be treated as vendor-reported results rather than independent proof that the models are equivalent across all tasks.
GLM-5.3-Flash Scores 57 On Artificial Analysis
Z.ai says GLM-5.3-Flash scored 57 on the Artificial Analysis Intelligence Index v4.1.1.
That places it in the same broad range as Claude Opus 4.8 on that index. Z.ai says the model achieves this at approximately $0.045 per task during the discounted period.
Intelligence And Cost
| Metric | GLM-5.3-Flash |
|---|---|
| Artificial Analysis Intelligence Index | 57 |
| Discounted cost per task | ~$0.045 |
| Total parameters | 320B |
| Active parameters | 18B |
The combination of capability and cost is central to Z.ai’s positioning.
GLM-5.3-Flash Costs About One-Tenth Of GLM-5.3
Z.ai says the new Flash model is priced at approximately one-tenth the price of GLM-5.3.
Standard API pricing is:
- $0.15 per million input tokens
- $0.50 per million output tokens
- $0.03 per million cached input tokens
API Pricing
| Model / Rate | Price Per 1M Tokens |
|---|---|
| GLM-5.3-Flash input | $0.15 |
| GLM-5.3-Flash output | $0.50 |
| GLM-5.3-Flash cached input | $0.03 |
| Limited-time discounted input | $0.075 |
| Limited-time discounted output | $0.25 |
A limited-time promotion cuts the standard input and output rates by 50%, according to current pricing information.
The Model Is Much Cheaper Than Claude Opus
The cost comparison becomes even more significant when GLM-5.3-Flash is compared with premium frontier models.
Claude Opus 4.8 is listed at $5 per million input tokens and $25 per million output tokens in the comparison cited by current reports.
API Price Comparison
| Model | Input / 1M Tokens | Output / 1M Tokens |
|---|---|---|
| GLM-5.3-Flash | $0.15 | $0.50 |
| Claude Opus 4.8 | $5.00 | $25.00 |
| GLM-5.3-Flash discount | $0.075 | $0.25 |
On standard list prices, GLM-5.3-Flash is roughly 33 times cheaper for input and 50 times cheaper for output than those Claude Opus 4.8 rates.
This is why Z.ai is emphasizing the model’s cost-efficiency rather than competing solely on benchmark scores.
The “One-Tenth Cost” Claim Needs Context
Z.ai’s one-tenth claim refers to its pricing relative to GLM-5.3.
It should not be interpreted as meaning that every real-world AI workload costs exactly 10% as much.
Actual costs can vary depending on:
- Input length
- Output length
- Cache utilization
- Hardware
- Throughput
- Context length
- Deployment configuration
- Task complexity
Similarly, a lower API price does not automatically mean lower total cost for every enterprise deployment.
Z.ai Uses Chinese-Made AI Chips
One of the most notable aspects of the release is that Z.ai says the model’s anonymous testing traffic was served entirely on Chinese-made AI chips.
According to reporting on the launch, Z.ai deployed more than 100,000 domestic chips across clusters supporting the model, although the company has not publicly identified the specific chip model.
That makes the model strategically significant beyond its pricing.
GLM-5.3-Flash Infrastructure
GLM-5.3-Flash
↓
Large-Scale AI Cluster
↓
Chinese-Made AI Chips
↓
Global Developer Traffic
↓
Real-World Performance TestingThe model therefore served real-world developer traffic rather than being evaluated only in a controlled laboratory environment.
Z.ai Optimized The Model For Domestic Hardware
Running a 320-billion-parameter model with a 1-million-token context window on domestic accelerators presents significant engineering challenges.
Z.ai says it modified the inference stack to address limitations in memory capacity, memory bandwidth and communication.
The company built a specialized inference engine based on SGLang and used techniques including:
- Tensor parallelism
- W8A8 quantization
- INT8/FP8/BF16 mixed cache quantization
- Layer Split
- Encode-Prefill-Decode architecture
According to Z.ai, these optimizations improved end-to-end service performance by 3x compared with its initial baseline on the same hardware.
Encode-Prefill-Decode Architecture
Z.ai also separates three major stages of inference:
- Multimodal encoding
- Prompt prefilling
- Token-by-token decoding
Each stage operates in a separate pool that can be independently scheduled and scaled.
Multimodal Input
↓
Encode Pool
↓
Prefill Pool
↓
Decode Pool
↓
OutputThis allows the infrastructure to allocate computing resources according to the requirements of each stage.
GLM-5.3-Flash Uses 30 Trillion Multimodal Training Tokens
Z.ai says the model was trained using a 30-trillion-token multimodal pre-training corpus.
The training data and architecture were designed to improve the model’s ability to handle text, images and video while maintaining lower computational requirements.
Training Scale
| Training Component | Figure |
|---|---|
| Multimodal pre-training corpus | 30T tokens |
| Total model parameters | 320B |
| Active parameters | 18B |
| Maximum context | 1M tokens |
The scale of the training corpus underscores that the Flash label does not refer to a small or lightly trained model.
GLM-5.3-Flash Is Built For AI Agents
Z.ai is positioning the model heavily around agentic workloads.
Its benchmark results include strong performance on:
- DeepSWE
- Terminal-Bench
- Toolathlon
- AutomationBench
- Agents’ Last Exam
- GDPval
These tests evaluate different aspects of coding, tool use and task completion.
Agentic Benchmarks
| Benchmark | GLM-5.3-Flash |
|---|---|
| Toolathlon Verified | 78.4 |
| AutomationBench v1.0.6 | 48.8 |
| Agents’ Last Exam | 26.3 |
| HLE With Tools | 55.3 |
| GDPval-AA v2 | 1,773 |
Z.ai says the model consistently outperforms GLM-5.2 across these evaluations.
Browser And Computer Use Are Built Into The Strategy
Z.ai says GLM-5.3-Flash’s multimodal capabilities can be used in ZCode for browser use and computer use.
An agent can visually inspect webpages, interact with them and operate desktop applications.
That is an important evolution from text-only coding models.
Code
↓
Run Application
↓
See Visual Output
↓
Understand Screen
↓
Modify Code
↓
Run Again
↓
Verify ResultThis feedback loop can be particularly useful for web development and software agents.
The Model Can Understand Visual Results
Z.ai says it developed synthetic training data for visual coding, allowing the model to view the output of code and make further changes based on visual feedback.
This can be useful when generating:
- Web interfaces
- Mobile applications
- 3D scenes
- Interactive visualizations
- Game environments
The objective is to move beyond generating code toward generating and verifying working software.
GLM-5.3-Flash Is Open Weight
Z.ai has made the model weights publicly available through Hugging Face.
The model can be deployed through inference frameworks including SGLang, vLLM and TokenSpeed, according to Z.ai.
This gives developers an alternative to using the hosted Z.ai API.
Deployment Options
| Option | Benefit |
|---|---|
| Z.ai API | Managed inference |
| Hugging Face weights | Self-hosting |
| SGLang | Inference optimization |
| vLLM | Popular serving framework |
| TokenSpeed | Alternative inference support |
| ZCode | Coding and agent workflows |
The model’s open availability could help Z.ai expand adoption among developers and researchers.
GLM-5.3-Flash Vs. Traditional “Flash” Models
Historically, models labeled “Flash,” “Mini” or “Air” have generally emphasized speed and lower cost at the expense of some capability.
Z.ai is trying to change that positioning.
GLM-5.3-Flash retains a 320B total parameter count, supports multimodal input and reaches frontier-model-level benchmark results while targeting much lower prices.
Traditional Flash
Smaller Model
↓
Faster + Cheaper
↓
Lower Capability
GLM-5.3-Flash
Large MoE Model
↓
Sparse Activation
+
Efficient Attention
+
Optimized Serving
↓
High Capability + Lower CostThe model is therefore “Flash” primarily because of its efficiency strategy rather than because it is a small model.
Z.ai Is Increasing Pressure On AI Pricing
The release arrives during an intense AI price competition among Chinese model developers.
Alibaba’s Qwen team has released Qwen3.8-Flash-Next, while DeepSeek and Moonshot AI have also been pushing high-capability models at relatively low prices.
GLM-5.3-Flash adds another powerful competitor.
China’s AI Model Competition
| Company | Recent Model Strategy |
|---|---|
| Z.ai | GLM-5.3-Flash |
| Alibaba | Qwen3.8-Flash-Next |
| DeepSeek | Low-cost open models |
| Moonshot AI | Kimi series |
| Huawei | Domestic AI compute ecosystem |
The competition increasingly revolves around three variables:
Capability + Cost + Compute Efficiency
The Bigger Picture
Z.ai’s GLM-5.3-Flash is significant because it challenges the assumption that frontier-level AI performance must come with frontier-level pricing. The model combines 320 billion total parameters with 18 billion active parameters, native text-image-video capabilities and a 1-million-token context window. Z.ai says it outperforms GLM-5.2 across benchmarks and real-world workloads while being priced at roughly one-tenth of GLM-5.3.
The model’s connection to the mysterious Ox Alpha release makes the launch even more notable. Z.ai secretly tested GLM-5.3-Flash on OpenRouter and OpenCode to gather real developer feedback before revealing its identity. The company says that traffic was served entirely on Chinese-made AI chips, adding another dimension to the release as Chinese AI companies increasingly demonstrate that domestic computing infrastructure can support large-scale model inference.
Looking Ahead
The biggest question will be whether GLM-5.3-Flash can sustain its reported performance when developers deploy it across a much wider range of real-world applications. Z.ai’s own benchmark results are encouraging, particularly in coding, agentic tasks and multimodal workloads, but independent testing will be important for assessing reliability, latency, long-context performance and total cost of ownership. The model’s open weights also give researchers and developers an opportunity to test those claims outside Z.ai’s infrastructure.
For the broader AI market, the release adds further pressure on premium model providers. A 320B-parameter multimodal system with 18B active parameters, 1M-token context and API pricing as low as $0.15 per million input tokens could make high-volume agentic applications substantially cheaper. If Z.ai can maintain that economics while delivering near-frontier performance, GLM-5.3-Flash could become an important example of how architectural efficiency, sparse computation and specialized inference infrastructure are reshaping the economics of advanced AI.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
