glm 5.3 flashx gguf, glm-5.3-flashx ollama, and the misspelling lcoal are all asking for a file that Z.ai did not publish. GLM-5.3-FlashX is a hosted serving tier. The weights you can download are GLM-5.3-Flash. Pulling those weights does not reproduce the hosted model code glm-5.3-flashx or the vendor's 200 tokens/s claim.
This article is based on the Z.ai Flash/FlashX overview, the GLM-5.3-Flash announcement, and the Hugging Face model card for zai-org/GLM-5.3-Flash, checked October 12, 2026. We did not find an official FlashX GGUF, an official Ollama library tag, or a repository named GLM-5.3-FlashX. Community quants may show up later. They are not the FlashX tier, and this page does not endorse one.
The direct answer
You can run GLM-5.3-Flash locally if you have the hardware and a framework the model card lists. You cannot download GLM-5.3-FlashX as its own checkpoint, because Z.ai documents FlashX as the high-speed variant of Flash: same overview page, separate API code, a stated inference speed of 200 tokens/s. Speed here is how the model is served, not a second weight release.
If the goal is that hosted tier, call glm-5.3-flashx on Z.ai's API. If the goal is a browser session of the standard Flash model, open glm5.app. The model id on this site is glm-5.3-flash. FlashX is the hosted speed tier of that same Flash model, not a second download.
What Hugging Face actually hosts
The official card is zai-org/GLM-5.3-Flash. A BF16 sibling is published as zai-org/GLM-5.3-Flash-BF16. A tree listing for the main repo shows the snapshot sharded into 62 safetensors files at about 226 GB. A separate public parameters note for GLM 5.3 Flash lists 328 GB for the weights. Those two figures are not the same file cut. Provision disks from the card you actually download, not from a single rounded number. Either way it is a datacenter download, not an Ollama one-liner on a laptop.
The model card describes the same Flash architecture the overview uses for this family: 320B total parameters, 18B active, hybrid sparse and linear attention, native multimodal input. The announcement says those weights are public and names SGLang, vLLM, and TokenSpeed for local inference, with more frameworks on the way. The model card README also points at Transformers, KTransformers, and Unsloth. None of those links are a FlashX GGUF.
Why the GGUF and Ollama queries exist
Searchers see "open weights" on the Flash announcement and attach the usual local formats: GGUF, Ollama, ModelScope typos (modescope), and "open source." Three mismatches follow.
- Format. The published files are safetensors shards for the frameworks above. We did not find an official Z.ai GGUF for Flash or FlashX on the model card.
- Name. Ollama-style tags people type (
glm-5.3-flashx) are not an official library entry we could verify. Existing Ollama guides on this site are about other GLM checkpoints, not this serving tier. - Size. A 226 GB tree listing or a 328 GB weights figure, plus KV cache for a 1M context window, is not a consumer-GPU pull. A quantized unofficial file that fits in 24 GB of VRAM is a different artifact. It will not match
glm-5.3-flashxquality or the 200 tokens/s hosting claim, and we are not going to pretend a random quant is the model.
"Open source" on the announcement means the Flash weights are published for you to deploy. It does not mean FlashX is a separate open checkpoint, and it does not mean Ollama can stream the fast tier to a desktop.
What a real local install is
Use the Flash weights with a framework the card or the announcement names. Expect to plan GPU memory for a 320B-total MoE, not for an 18B dense model. The "18B active" figure is compute per token, not the size of the download. You still load the expert pool.
After it runs, you have GLM-5.3-Flash on your hardware. Throughput is then your batch size, your kernels, and your GPUs. Z.ai's 200 tokens/s sentence describes the hosted FlashX tier on their stack. Copying the weights does not copy that stack. The announcement's own serving write-up is about a large cluster, disaggregated encode/prefill/decode, and kernel work. That is the opposite of ollama pull.
If you need the fast tier's behavior more than you need the files, stay on the API. As of the October 12, 2026 docs check, FlashX is an API model code, not a Coding Plan model. The overview's plan section still says FlashX is not yet available there, while glm-5.3-flash is.
A practical split
| You need | Do this |
|---|---|
The hosted fast tier (glm-5.3-flashx, vendor 200 tokens/s claim) | Z.ai API. There is no weight file for this row |
| The published weights on hardware you control | zai-org/GLM-5.3-Flash via SGLang, vLLM, TokenSpeed, or another framework the card lists |
| A small GGUF in Ollama | A different, smaller model. Do not label it FlashX |
| GLM 5.3 Flash without hosting 226 GB | glm5.app and model id glm-5.3-flash |
FAQ
Is there a GLM-5.3-FlashX GGUF?
Not from Z.ai, on the model card we checked. The published artifacts are GLM-5.3-Flash safetensors (and a BF16 repo). A third-party GGUF would be a conversion of Flash weights at best, not an official FlashX release.
Can Ollama run GLM 5.3 FlashX?
We could not verify an official Ollama tag for glm-5.3-flashx. Even a future tag for the Flash weights would still be Flash, and the 226 GB snapshot is far beyond a typical desktop Ollama install.
If I deploy GLM-5.3-Flash, did I deploy FlashX?
No. You deployed the weights. FlashX is the hosted high-speed serving tier and the API code glm-5.3-flashx. Local tokens per second are yours to measure. They are not the vendor's 200 tokens/s line.
Where do people get the "open source" wording?
From the GLM-5.3-Flash announcement, which says the weights are public and lists inference frameworks. Read that as Flash, then check the overview again before you budget a FlashX local box.
Sources
- GLM-5.3-Flash/FlashX overview — FlashX described as the live high-speed tier at 200 tokens/s, with model code
glm-5.3-flashx. Checked October 12, 2026. - GLM-5.3-Flash announcement — public weights, and SGLang, vLLM, and TokenSpeed named for local deployment.
- zai-org/GLM-5.3-Flash — official weights, 62-shard safetensors snapshot listed around 226 GB, and the framework list on the model card. No FlashX repo and no official GGUF were on that card.
- zai-org/GLM-5.3-Flash-BF16 — the BF16 weight repo, still Flash, not FlashX.
