GLM-5.3: Z.ai's 753B Open Model — Sizes, Prices, How to Use It
GLM-5.3 is Z.ai's 753B open-weight flagship, ungated on Hugging Face. We cover GLM-5.3-Flash, hosted prices from $0.50/M tokens, local limits, and agent use.

GLM-5.3 is the new open-weight flagship from Z.ai, the company Chinese developers know as Zhipu AI (智谱). It has 753 billion parameters, which makes it far too big for your laptop — and you can download it anyway. The weights sit on Hugging Face with no gate: no access form, no approval queue, no "describe your use case" box.
It shipped with a smaller sibling, GLM-5.3-Flash: 321 billion parameters, multimodal (it reads images as well as text), and listed at about $0.50 per million tokens on third-party hosts. Early community testing has centered on coding and agentic tasks, which is also where most people will put it to work.
Below: what's in the release, what it costs today, why running it at home is mostly a fantasy, and who should bother.
What GLM-5.3 is, and why ungated weights matter
Per Z.ai's official zai-org page on Hugging Face, GLM-5.3 is a text-generation model with 753,329,940,480 parameters. The repo was last updated on October 14, 2026, so this is a release only weeks old.
The size is impressive, but the missing gate is the bigger story. Plenty of "open" models make you request access and wait. This one doesn't, and that has three practical consequences:
- Providers compete to serve it. Anyone can mirror and host the weights. By early November, Hugging Face's provider listing showed at least six third-party hosts for the flagship, at prices from $4.00 to $6.27 per million tokens. When the model is identical everywhere, price and speed become the whole pitch.
- You can switch hosts without switching models. If your provider raises prices or has an outage, you move an API key, not your prompts.
- You can pin a version. An API-only model can be retired on someone else's schedule. Weights on your own disk keep working until you decide otherwise.
One caveat: ungated means no approval step, not no license terms. Read the license on the model card before you build a commercial product on it.
The model family at a glance
The family on Hugging Face comes down to two models, each with a standard repo and a full-precision BF16 copy.
| Model | Parameters | Input and output | Hugging Face repo | Best use |
|---|---|---|---|---|
| GLM-5.3 | 753B | Text in, text out | zai-org/GLM-5.3 | The hardest coding and agent tasks |
| GLM-5.3-BF16 | 753B | Text in, text out | zai-org/GLM-5.3-BF16 | Quantizing or fine-tuning from full precision |
| GLM-5.3-Flash | 321B | Images and text in, text out | zai-org/GLM-5.3-Flash | Daily work, screenshots, cost-sensitive loops |
| GLM-5.3-Flash-BF16 | 321B | Images and text in, text out | zai-org/GLM-5.3-Flash-BF16 | Full-precision base for Flash derivatives |
Flash has 321,323,031,390 parameters, and its Hugging Face task tag is image-text-to-text: hand it a screenshot, a chart, or a photo of a whiteboard along with your prompt, and it answers in text. The flagship is text-only. If your workflow involves images, that settles the choice before price even comes up.
The BF16 repos are the full-precision checkpoints, the right starting point if you plan to quantize or fine-tune yourself. If you're calling a hosted API, the provider picks the precision for you — one reason the same model can behave slightly differently from one host to the next.
What GLM-5.3 costs to use today
Z.ai runs its own API, but the easiest comparison is Hugging Face's inference-provider listing, which puts third-party hosts side by side. As of early November 2026:
| Model | Provider | Listed price per 1M tokens |
|---|---|---|
| GLM-5.3 | DeepInfra | $4.00 (lowest) |
| GLM-5.3 | Together AI, Nscale, Novita, Baseten | About $4.40 |
| GLM-5.3 | Scaleway | $6.27 |
| GLM-5.3-Flash | Listed providers, including Together AI | About $0.50 |
Two caveats before you budget. These are third-party hosting prices; Z.ai's own API pricing may differ. And many providers bill input and output tokens at different rates, so check the full rate card on the provider's site. The listing is a starting point, not an invoice.
At the listed rates, an agent session that burns 2 million tokens reading a repo, planning, and iterating costs roughly $8-$9 on the flagship's cheaper hosts and about $1 on Flash. One-eighth of the price changes how freely you let an agent loop.
Why you probably won't run GLM-5.3 locally
Open weights invite the obvious question: can I just run it myself? For the flagship, the honest answer is no — not on anything you'd call a workstation.
The math is unforgiving. In BF16, 753 billion parameters is about 1.5 TB of weights, before a single byte of KV cache for context. Squeeze it to 8 bits per weight and you're still near 750 GB; at 4 bits, around 380 GB. For scale, a server with eight 80 GB GPUs has 640 GB of GPU memory, which can't hold the flagship even at 8-bit.
Flash is friendlier, but not friendly: 321B parameters works out to roughly 640 GB in BF16, 320 GB at 8-bit, and 160 GB at 4-bit, again before cache.
So the realistic options are:
- Hosted providers. For almost everyone, this is the answer. Pay per token and let someone else own the GPUs.
- The Flash tier. If you must self-host for data-control reasons, Flash is the more tractable target: still multi-GPU territory, but less than half the flagship's footprint.
- Quantized multi-GPU clusters. Labs and companies with their own hardware can spread a quantized flagship across several high-memory GPUs. If that's you, you already ran these numbers.
If someone online says they're running GLM-5.3 on a laptop, they're running a different model — or measuring speed in tokens per minute.
GLM-5.3-Flash: the one most people should start with
The flagship gets the headlines; GLM-5.3-Flash will probably do most of the work. Three reasons:
- It sees. Multimodal input means you can paste a screenshot of a broken layout, an error dialog, or a chart and ask about it directly, instead of describing pixels in words.
- It's cheap enough to loop. At about $0.50 per million tokens, an agent can retry, re-read files, and run long sessions without the bill becoming the main character.
- It's fast where it counts. Together AI serves it at about 272 tokens per second, quick enough that interactive coding never feels like waiting for a fax.
The usual routing pattern applies: send routine edits, screenshot questions, and high-volume jobs to Flash, and escalate to the 753B flagship when a task needs deeper reasoning across a large codebase. And don't read "Flash" as "toy" — at 321B parameters, it's a large model in its own right.
Using it as a coding agent's brain
Since early testing points at coding and agentic work, its natural home is a bring-your-own-model coding agent. OpenCode is the obvious candidate: open source, terminal-first, and able to connect to 75+ model providers. Kimi Code is another terminal agent with configurable providers if you want a second option.
A setup that works:
- Pick a host from the price table and create an API key.
- Add that provider in your agent. If it isn't in the built-in list, register its OpenAI-compatible endpoint as a custom provider; most of the hosts above offer one.
- Use the flagship for planning-heavy sessions and big refactors, and switch to Flash for quick fixes and anything involving screenshots.
- Watch token usage for the first few sessions. Agents are much hungrier than chat, and per-token pricing makes that visible fast.
Agentic work also means tool calls. If you're wiring up MCP servers — say REA, the reverse-engineering MCP we covered, which hands an agent a large menu of analysis tools — the model's tool-calling discipline matters as much as its raw coding skill. Test with your real tool stack before you move a whole team over.
Open weights in context: GLM-5.3, Kimi K3.1 Pro, and DeepSeek
Z.ai's release lands in a crowded autumn for open models, and the useful distinction right now is between weights you can download today and weights that are promised.
- GLM-5.3: public and ungated on Hugging Face now, with both a flagship and a multimodal Flash tier.
- Kimi K3.1 Pro: Moonshot has announced open weights, scheduled for November 4. Until they land, you're using it through an API. Our Kimi K3.1 breakdown covers what's known.
- DeepSeek: if cost per token matters more to you than downloadable weights, read our notes on DeepSeek 4.1 Flash before committing, or start from the DeepSeek tool page.
For teams that care where their data goes, downloadable weights are the point: start on a hosted API and move to your own infrastructure later without changing models. For everyone else, they're the reason half a dozen hosts are undercutting each other on price.
Who should pick it, and which tier
Pick the 753B flagship if you want an open-weight model for hard coding and agent work, you're comfortable paying around $4 per million tokens, and you value the freedom to switch hosts or self-host later.
Pick GLM-5.3-Flash if you need image input, you run lots of agent iterations, or cost per task matters more than the last bit of capability.
Look elsewhere if you want a polished subscription product instead of an API you wire up yourself — a first-party agent like Claude Code is the less fiddly path. The same goes if you need independent benchmarks before you commit: early community testing on coding is encouraging, but it isn't a rigorous eval on your own workload. And if the plan was to run it on your own machine, Flash included, adjust the plan.
FAQ
Is GLM-5.3 open source?
The weights are public and ungated on Hugging Face, so anyone can download them without approval. "Open weights" isn't automatically an open-source license, though — check the license on the model card for commercial terms.
How much does GLM-5.3 cost?
Third-party hosts on Hugging Face list the flagship at about $4.00 to $6.27 per million tokens, with DeepInfra lowest at $4.00, and Flash at about $0.50. Z.ai's own API pricing may differ, and input and output tokens are often billed separately.
Can I run GLM-5.3 locally?
Not on a normal computer. The BF16 weights alone are roughly 1.5 TB before KV cache, and even a 4-bit quantization needs around 380 GB. For individuals, hosted providers are the realistic route; self-hosting is a lab-scale, multi-GPU project.
What's the difference between GLM-5.3 and GLM-5.3-Flash?
GLM-5.3 is the 753B text-only flagship. GLM-5.3-Flash is a 321B multimodal model that accepts images and text, at roughly an eighth of the hosted price. Start with Flash and escalate to the flagship for the hardest tasks.
Can I use GLM-5.3 with coding agents like OpenCode?
Yes. Any bring-your-own-model agent that accepts an OpenAI-compatible endpoint can use it through a hosted provider, and OpenCode is a good place to start: add a provider key, then select the model.


