GLM-5.3-Flash: 320B MoE, 18B Active, MIT

There is one race in open-weight models this summer: producing the same intelligence with less compute. Alibaba’s Qwen3.8-Flash-Next beat its own 397B sibling with 6 billion active parameters. Z.ai’s answer is GLM-5.3-Flash: 320 billion total parameters with only 18 billion running per token, leaving GLM-5.2 behind at a tenth of the cost.

Two details separate this from an ordinary “cheap tier” announcement. The first is the MIT licence. The second is a line near the end of the announcement: all of that traffic ran not on NVIDIA, but on domestically developed Chinese accelerators.

Let’s take it in order. 👇🏻

What Is GLM-5.3-Flash?

A mixture-of-experts (MoE) model with downloadable weights that reads text and images and writes text. It is the first natively multimodal member of the GLM-5 series, meaning vision is not a bolt-on adapter but part of the base model’s training.

The middle row is the one that matters. In an MoE model the total parameter count is a marketing number; what drives your bill and your latency is how many parameters fire per token. Here it is about 5.6% of 320 billion.

For comparison, Z.ai’s own GLM-4.5: 355 billion total, 32 billion active, 92 layers. GLM-5.3-Flash nearly halves both the active parameter count and the layer count at a similar total size. Halving layers matters for latency directly: each token passes through fewer sequential steps.

What Changed in the Architecture? 🧠

Z.ai describes changes in three areas: attention, residual connections and the pre-training corpus.

1. Hybrid attention: linear plus sparse

Classic full attention compares every new token against every past token. As context grows, both compute and the KV cache blow up. For the first time in the GLM series, that job is split in two:

• Linear attention compresses history into a fixed-size state and carries local dependencies. It remembers.
• Sparse attention uses a lightweight indexer to find the relevant regions of the context and pull exact information from them. It retrieves.

That split is not new; DeepSeek and Qwen do something similar. Z.ai’s own contribution is this: at a 1M-token context the indexer itself becomes a meaningful latency and memory cost. The fix is IndexPool, which compresses four indexer key vectors into one through weighted pooling. The cost of deciding where to look is divided by four.

The result, by Z.ai’s own measurement: 3x less attention compute and a 4.4x smaller KV cache than GLM-5.3. Among the models compared, GLM-5.3-Flash has the lowest attention compute of all. On KV cache size it still sits slightly behind Kimi-K3 and DeepSeek-V4-Flash, which Z.ai explicitly flags as room for improvement.

2. mHC: residual connections for scaling efficiency

The model also adopts Manifold-Constrained Hyper-Connections (mHC). In a normal Transformer every layer writes to and reads from the same residual stream, so as the network deepens, early information gets diluted by everything written after it. Hyper-connection approaches widen that single lane into several and make each layer’s read and write weights learnable. The “manifold-constrained” part keeps that extra freedom inside a geometric constraint so training stays stable.

In Z.ai’s words the payoff is improved scaling efficiency: more capacity from the same compute.

3. A 30 trillion token multimodal corpus

The third leg is data. Pre-training used a 30 trillion token corpus that was multimodal from the start. Because vision is not pasted on afterwards, the model can also decide when to look at an image, which is the prerequisite for inspecting a screenshot of its own output while it writes code.

How does the base model look?

Scores from the raw, pre-post-training model are the most honest way to see whether the architecture actually paid off:

Read it like this: GLM-5.3-Flash-Base beats GLM-4.5-Base overall with fewer active parameters, and stays close to the 40B-active / 744B-total GLM-5-Base on most benchmarks. On code (LiveCodeBench-Base 37.6) it clears both by a visible margin. Trailing GLM-5-Base on SimpleQA is the expected outcome: 744 billion total parameters simply store more facts.

Vision Inside the Coding Loop 👁️

This is the part Z.ai leans on hardest. For frontend work, game development and 3D simulation, the final output is not code; it is an interface, an interaction, something a user sees. Most failures in that kind of work only surface once the page renders or the game is played. Tests green, screen broken.

So the model is trained to look at its own output: it interacts with the environment, inspects the interface it produced as an image, and rewrites the code based on what it sees. For frontend coding Z.ai also explored reinforcement learning with environment feedback and strengthened GUI judgment through agent-based verification grounded in real user flows. Validation moves from “is the function correct” to “is the rendered product correct”.

The same capability extends past code into office work: documents, spreadsheets, presentations, dashboards, meeting artifacts. Because the model can read those directly instead of waiting for you to transcribe them, it can evaluate its own output even on subjective qualities like presentation quality. The OfficeQA Pro and Chartography scores below measure exactly that.

Benchmark Scores 📊

All numbers come from Z.ai’s own evaluations. Columns: GLM-5.3-Flash, GLM-5.2, DeepSeek-V4-Vision-Exp, Claude Opus 4.8, GPT-5.6 Terra and Gemini 3.7 Flash.

Coding:

Agentic:

Vision:

What stands out: the jump over GLM-5.2 is close to 2x on some tests (48.8 against 26.2 on AutomationBench, 63.4 against 46.2 on DeepSWE). On chart and document reading (Chartography 78.0, OfficeQA Pro 62.4) it clears every closed model in the table. On general visual reasoning (BabyVision) and video understanding (MVBench, MMVU), however, Gemini 3.7 Flash leads comfortably.

Z.ai also reports an internal number: on Z.ai Code Bench v1.0, run inside Claude Code 2.1.207, it scores 29.0 at max effort against Claude Opus 4.8’s 29.5.

On the Artificial Analysis Intelligence Index v4.1.1 it scores 57 at $0.045 per task with the discount applied. That level of intelligence has until now only been available at roughly 10x the cost.

ox-alpha: Anonymous Testing Before Launch

Before announcing anything, Z.ai put the model on OpenCode and OpenRouter under the codename ox-alpha to gather user feedback. Unbadged releases have become common this summer, and the logic is straightforward: see how a model behaves on real workloads with the brand halo removed.

By Z.ai’s account, ox-alpha became the most popular model of the week. The more interesting detail follows: all of that traffic was served on Chinese AI chips.

Serving at Scale on Chinese Chips: Why It Matters ⚙️

This section is more technical than a model announcement usually gets, and it is probably the most consequential part of the release.

Z.ai served GLM-5.3-Flash for a week on a large cluster of domestically developed accelerators connected by a high-bandwidth interconnect. Those chips are constrained mainly by memory capacity and bandwidth, especially at a 1M-token context. The techniques used:

• Intra-node tensor parallelism for linear attention and the LM head
• ReplaySSM
• W8A8 quantization with hybrid INT8 / FP8 / BF16 cache quantization
• Layer Split
• At cluster scale, an EPD (Encode-Prefill-Decode) disaggregated architecture: multimodal encoding, prompt prefill and token-by-token decoding split into independently scheduled worker pools

They built a dedicated inference engine for the hardware on top of SGLang. The most circular detail in the announcement lives here: the work was accelerated by their own GLM-5.3-powered infrastructure agent, which helped engineers write kernels and diagnose bottlenecks in the very system that serves the model.

The result: a 3x end-to-end serving improvement over their initial baseline on the same hardware, and per-token cost comparable to mainstream NVIDIA GPUs.

What Does It Cost? 💸

The discount runs until 9 September 2026 at 24:00 (UTC+8). Cached input storage is currently free.

For comparison, the same provider’s GLM-5.3 and GLM-5.2 both sit at $1.4 in / $4.4 out. That makes GLM-5.3-Flash roughly 9x cheaper on both sides. Z.ai’s “one tenth the price” framing uses the discounted rate.

It has also rolled out to GLM Coding Plan subscribers, where it grants 3x the usable quota of GLM-5.3.

You can put it side by side with Claude and GPT pricing for your own workload in our LLM cost calculator, and measure how many tokens your input actually costs with our token counter.

How Do You Use It?

The API is OpenAI-compatible:

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.environ.get("ZAI_API_KEY"),
    base_url="https://api.z.ai/api/paas/v4",
)

response = client.chat.completions.create(
    model="glm-5.3-flash",
    messages=[{
        "role": "user",
        "content": "Find the alignment bug in this screenshot and write the CSS fix.",
    }],
    stream=True,
)

With Claude Code: Z.ai also exposes an Anthropic-compatible endpoint, so two environment variables are enough:

export ANTHROPIC_BASE_URL=https://api.z.ai/api/anthropic
export ANTHROPIC_AUTH_TOKEN=<your_api_key>
claude

In Z.ai’s own agent, ZCode, the multimodal side unlocks through Browser Use and Computer Use: the agent clicks through web pages and visually verifies its own work, and it can drive desktop applications.

On your own server

The weights live on Hugging Face under zai-org/GLM-5.3-Flash. Supported inference engines: SGLang, vLLM, TokenSpeed and KTransformers.

vllm serve zai-org/GLM-5.3-Flash

SGLang:

python3 -m sglang.launch_server \
    --model-path zai-org/GLM-5.3-Flash \
    --host 0.0.0.0 --port 30000

Hugging Face already lists 12 quantizations and 2 finetunes, so community conversions for llama.cpp / Ollama / LM Studio have started appearing.

Who Is It For?

• High-volume coding agents: the price, the 1M-token context and the 78.4 on Toolathlon together are the model’s entire reason to exist. If you are on GLM-5.2 the switch is close to a no-brainer: better scores at a ninth of the cost.
• Frontend and interface work: a model that can look at a screenshot and correct its own output is a direct win for bugs that only appear once something renders.
• Document and table reading: the OfficeQA Pro and Chartography scores clear the closed models in the table. Worth a look for PDF and dashboard pipelines.
• Products that need an MIT licence: if you want a strong open-weight model with no fine print to embed in a commercial product, this is the easiest option of the summer.
• Video understanding: not this model. Gemini 3.7 Flash leads MVBench and MMVU by a clear margin.
• Single-GPU local inference: also not this model. All 320 billion parameters have to fit in memory.

Frequently Asked Questions

Q: When was GLM-5.3-Flash released? A: 26 August 2026, on the Z.ai API platform and Hugging Face (zai-org/GLM-5.3-Flash) the same day.

Q: How many parameters does GLM-5.3-Flash have? A: 320 billion total, with 18 billion active per token, across 45 layers of MoE. For comparison, GLM-4.5 was 355 billion total, 32 billion active and 92 layers.

Q: Is GLM-5.3-Flash open source? Can I download it? A: The weights are downloadable and the licence is MIT, which covers commercial use. The precise term is still “open weight”: the model files were released, the training data was not.

Q: What does GLM-5.3-Flash cost? A: $0.15 per million input tokens and $0.50 per million output tokens, or $0.075 and $0.25 with the 50% discount running until 9 September 2026. Cached input is $0.03. Self-hosting removes the API fee and adds hardware cost.

Q: How large is the context window? A: 1 million tokens. The hybrid linear plus sparse attention and IndexPool exist precisely to make that length economical to serve.

Q: What is the difference between GLM-5.3-Flash and GLM-5.3? A: Flash is built for efficiency: 3x less attention compute, a 4.4x smaller KV cache and roughly a ninth of the price. You give up some of GLM-5.3’s raw capability in exchange, but it still clears GLM-5.2 by a wide margin on coding and agentic tests.

Q: Does it take image input? A: Yes. It is the first natively multimodal member of the GLM-5 series, handling images and video frames in the same model as text, which is what lets it inspect a screenshot of the interface it just built.

Q: How much VRAM does it need? A: Roughly 320 GB in FP8, roughly 170-180 GB at 4-bit, plus the KV cache. Not a single consumer card. KTransformers support makes CPU offloading viable.

Q: Does it work with Ollama or llama.cpp? A: The official list is SGLang, vLLM, TokenSpeed and KTransformers. Community quantizations have started landing on Hugging Face, so llama.cpp and Ollama conversions exist. It is a new architecture, so use current versions of those tools.

Q: Can I use it with Claude Code? A: Yes. Set ANTHROPIC_BASE_URL to https://api.z.ai/api/anthropic and pass your key as ANTHROPIC_AUTH_TOKEN, and Claude Code talks to the model directly. Z.ai’s GLM Coding Plan subscription covers this usage too.

Q: What was ox-alpha? A: The codename GLM-5.3-Flash was tested under anonymously on OpenCode and OpenRouter before launch, to collect feedback on real workloads without the brand halo. Z.ai says it became the most used model of that week.

Take care… 🙂