MITmultimodalRuns locally320B (card)

GLM-5.3-Flash: Specs, License & How to Run

MIT multimodal MoE (card 320B/18B, 1M context) from Z.ai’s GLM-5 series.

Official GLM-5.3-Flash benchmark figure from the zai-org/GLM-5 repository — Official project image. Source · License: MIT (weight card) / Apache-2.0 (GLM-5 tree).

License: MIT
Parameters: 320B (card) (active 18B (card))
HF likes: 2,941 (2026-10-25)
First seen: 2026-08-25

TL;DR. GLM-5.3-Flash is a multimodal model from Z.ai. Parameters: 320B (card) (active 18B (card)).License: MIT. It is listed under ChatGPT, Closed LLM APIs.

Developer	Z.ai
Type	multimodal
Architecture	Native multimodal MoE (HF model_type glm5_next; architecture Glm5NextForConditionalGeneration). Card: hybrid sparse + linear attention and Manifold-Constrained Hyper-Connections (mHC). text_config: 45 layers, hidden 4096, 288 routed experts, 8 active, 1 shared, max_position_embeddings 1,048,576. Vision: 24-layer glm5_next_vision. safetensors.parameters sum to 321,323,031,390.
Parameters	320B (card) · active 18B (card)
Modalities	text, image, video
Languages	en, zh
Context	1,048,576 tokens
License	MIT — LICENSE on the Hugging Face card is the MIT License, copyright 2026 Z.AI Co., Ltd. cardData.license is mit. The companion GitHub tree zai-org/GLM-5 is Apache-2.0 (code/docs, not a substitute for the weight LICENSE). Not legal advice.
Hardware	Not yet verified — Card usedStorage is 656,694,929,761 bytes (~612 GiB). Peak VRAM is not independently measured. README points at SGLang, vLLM, TokenSpeed, Transformers, KTransformers, and Unsloth. (not independently tested)
Hugging Face likes	2,941 (as of 2026-10-25)
Model card	https://huggingface.co/zai-org/GLM-5.3-Flash
First seen	2026-08-25

GLM-5.3-Flash is a natively multimodal Mixture-of-Experts model from Z.ai. Official hosts we opened on 2026-10-25: the ungated card zai-org/GLM-5.3-Flash (created 2026-08-25T06:43:14Z; last modified 2026-10-15T06:32:52Z; 2,941 likes / 6,377,661 downloads), the MIT LICENSE, the blog GLM-5.3-Flash, and the companion tree zai-org/GLM-5 (7,383 stars, 998 forks, Apache-2.0, last push 2026-10-15T01:36:08Z).

This page hangs off ChatGPT alternatives and closed LLM APIs. That mapping is editorial. We have no sourced quality comparison against ChatGPT. Head-to-head facts versus the MIT 1017 checkpoint already on this site: GLM-5.3-Flash vs DeepSeek-V4.1-Flash.

The card calls it the first natively multimodal model in the GLM-5 series. Inputs include text, image, and video; Hugging Face pipeline_tag is image-text-to-text. Vendor sentences about beating GLM-5.2 or approaching Claude Opus 4.8 stay vendor.

What we verified (2026-10-25)

Field	Value	Source
Host	huggingface.co/zai-org/GLM-5.3-Flash (not gated)	HF API
Created / last modified	2026-08-25T06:43:14Z / 2026-10-15T06:32:52Z	same
Likes / downloads	2,941 / 6,377,661	same
License tag + file	mit · MIT License (© 2026 Z.AI Co., Ltd)	HF `cardData` + LICENSE
Pipeline tag	image-text-to-text	HF API
Architecture	glm5_next (Glm5NextForConditionalGeneration)	card + HF config
Params (card)	320B total / 18B activated	card
safetensors.parameters	321,323,031,390	HF API
usedStorage	656,694,929,761 bytes (~612 GiB)	same
Context	1,048,576 tokens (`text_config.max_position_embeddings`)	HF config
Experts (config)	288 routed, 8 active, 1 shared; 45 text layers	same
Companion GitHub	zai-org/GLM-5, 7,383 stars, Apache-2.0	GitHub API

Hugging Face’s UI can show a ~321B “model size” line from the safetensors sum. We keep the README’s 320B / 18B quote and do not invent a reconciliation — same pattern as DeepSeek-V4.1-Flash and MiniMax M3.1 Flash.

How to run (from the card, not tested)

Recommended serving frameworks copied from the README: SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth.

README notes we copied:

reasoning_effort accepts low, high, and max (default max if omitted). The card says to keep max for benchmark reproduction.
In the chat template, clear_thinking defaults to false; the card says to pass clear_thinking=true for chat.

Hosted docs.z.ai is an API SKU. forkalt has not served the shards.

License

LICENSE is the MIT License, copyright 2026 Z.AI Co., Ltd. The README YAML and cardData.license also say mit. The GitHub tree zai-org/GLM-5 is Apache-2.0 and is not a substitute for the weight LICENSE. Not legal advice.

What we did not invent

VRAM, tokens/s, and “one-tenth the price” stay vendor.
The card’s bench_53.png figure is identification, not a forkalt table. We do not OCR scores out of that PNG.
We do not rank GLM-5.3-Flash against ChatGPT or DeepSeek-V4.1-Flash.

Related: GLM-5.3-Flash vs DeepSeek-V4.1-Flash, DeepSeek-V4.1-Flash, MiniMax M3.1 Flash, MiMo-V2.6-Flash, Kimi K3, vLLM, DeepInfer vs vLLM. Self-hosted UIs: Open WebUI, LibreChat, Jan. Pixel-editor traffic on this site still lands on PhotoCraft vs GIMP — a different job.

Deploy options

Comparisons

GLM-5.3-Flash vs DeepSeek-V4.1-Flash

FAQ

Are the GLM-5.3-Flash weights actually public?

Yes on the host we opened. zai-org/GLM-5.3-Flash is ungated (created 2026-08-25T06:43:14Z; last modified 2026-10-15T06:32:52Z; 2,941 likes / 6,377,661 downloads in the 2026-10-25 API check). MIT LICENSE is in the same repo. Hosted Z.ai API SKUs are a different artifact.

Is this MIT?

Does it replace ChatGPT or beat DeepSeek-V4.1-Flash-1017?

No sourced bake-off on this site. Vendor adjectives such as “outperforms GLM-5.2” or “approaching Claude Opus 4.8” stay vendor. Head-to-head facts we can copy: GLM-5.3-Flash vs DeepSeek-V4.1-Flash.

How big is it?

Card: 320B total / 18B active, 1M context. Hugging Face safetensors.parameters sum to 321,323,031,390. usedStorage 656,694,929,761 bytes. We keep the README 320B/18B quote and do not invent a reconciliation.

Sources & verification

Last verified 2026-10-25 · Last updated 2026-10-25· Edited by forkalt editors · Editorial policy

License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.