GLM-5.3-Flash

Updated
26.08.2026
Thinking
Tools
Vision
Reasoning
Code
Multilingual
Run

GLM-5.3-Flash is a 320B MoE with 18B active parameters, the first natively multimodal model in the GLM-5 series, released under MIT.

At a glance

  • License: MIT
  • Parameters: 320B total, 18B active per token
  • Context length: 1,048,576 tokens (1M), per config.json
  • Modalities: Text and image input
  • Minimum hardware: 328.3 GB for the FP8 weights, 642.7 GB for BF16

What is GLM-5.3-Flash?

GLM-5.3-Flash is a 320B-parameter Mixture-of-Experts model from Z.ai (zai-org) and the first natively multimodal model in the GLM-5 series. Just 18B parameters are active per token, so the compute spent on each token is closer to a small model while the weight set stays very large. Z.ai published it on August 25, 2026 under the MIT license.

SpecificationGLM-5.3-Flash
Total parameters320B (the repo metadata reports about 321B across all weights)
Active parameters18B per token
Architectureglm5_next, sparse MoE with hybrid sparse and linear attention
Layers45
Experts288 routed, 8 active per token
Hidden size4096
Context window1,048,576 tokens (1M), per config.json
ModalitiesText and image input, with image and video token ids in the config
Pre-training corpus30T multimodal tokens
Default weightsFP8
Release dateAugust 25, 2026
LicenseMIT

The architecture is registered as glm5_next. It combines sparse and linear attention in one hybrid stack, which Z.ai calls a first for the GLM series and says sharply reduces long-context serving costs while preserving precise long-context capabilities. It also adopts Manifold-Constrained Hyper-Connections (mHC) to improve scaling efficiency. The 45 layers route every token to 8 of 288 experts, and config.json carries both a text_config and a vision_config alongside dedicated image and video token ids, so the vision path is part of the base model rather than an adapter bolted on later. One number to read carefully: the 1M context comes from max_position_embeddings in config.json. The README itself never states a context length.

What GLM-5.3-Flash is good at

There is no benchmark table to show you here, and that is a fact about the release rather than an omission on our side. Z.ai published the results for this model as an image in the model card and put no scores in text, so there are no numbers to reprint or check. What the card does say in words is the claim itself. According to Z.ai, GLM-5.3-Flash "outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks". Both halves of that are the vendor's own framing, measured by the vendor, against GLM-5.2.

The footnotes name the benchmarks that were run: HLE with tools, NL2Repo, DeepSWE, Terminal-Bench 2.1, Agent's Last Exam, Toolathlon Verified, AutomationBench v1.0.6, GDPval-AA v2 and BabyVision. Z.ai does not describe what any of them measure, so treat the names as names. The footnotes do give the harness settings, and those are worth reading: DeepSWE was run under 400K context in the mini-swe-agent harness with a 6 hour timeout, NL2Repo under 1M context with 64k new tokens, HLE with tools at 300K context with a context-management strategy and GPT-5.6-luna as judge, and Terminal-Bench 2.1 inside Claude Code 2.1.207. Toolathlon Verified is pass@1 averaged over three runs through the official evaluation service, and GDPval-AA v2 was scored by Artificial Analysis. If you want to reproduce any of it on your own hardware, those are the settings to match.

GLM-5.3-Flash hardware requirements

The system requirement to check is memory, and the only figures this release gives are file sizes: no GGUF builds have been published yet, so the weights you can download today are the two safetensors repos.

MemoryBuild to pickFile size
328.3 GB or morezai-org/GLM-5.3-Flash, FP8, 62 shards328.3 GB
642.7 GB or morezai-org/GLM-5.3-Flash-BF16, 120 shards642.7 GB

When both fit, take the FP8 repo: it is the default Z.ai ships, and it is half the download. For serving those weights the model card names four stacks, SGLang, vLLM, TokenSpeed and KTransformers, each with its own recipe page. A quantized build will be smaller than both rows above, and if that route is new to you, start with what GGUF is.

How to run GLM-5.3-Flash in Atomic Chat

Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.

  1. Download Atomic Chat for your platform and open it.
  2. Search for GLM-5.3-Flash in the model browser and open Download Options.
  3. Pick the build that fits the memory you have, then start a chat.

One caveat on the third step: Atomic Chat will publish the GGUF builds for this model as AtomicChat/GLM-5.3-Flash-GGUF, and that repo holds no .gguf files yet, so there is nothing to pick until they land. For the rest of the lineup, see every GLM model you can run locally.

GLM-5.3-Flash license

GLM-5.3-Flash is released under MIT. That is about as permissive as open weights get: commercial use, modification, redistribution and private forks are all allowed, with no revenue threshold, no acceptable-use appendix and no obligation beyond keeping the copyright notice. It is also what lets anyone requantize these weights and republish them without asking permission first.

Get the weights from Hugging Face

huggingface-cli download zai-org/GLM-5.3-Flash
# full-precision weights, 642.7 GB instead of 328.3 GB:
huggingface-cli download zai-org/GLM-5.3-Flash-BF16
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org/GLM-5.3-Flash",
    "messages": [{"role": "user", "content": "Refactor this module and explain the change."}]
  }'
# Z.ai names SGLang, vLLM, TokenSpeed and KTransformers for serving the original weights.
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="local")
res = client.chat.completions.create(
    model="zai-org/GLM-5.3-Flash",
    messages=[{"role": "user", "content": "Summarise this repository layout."}],
)
print(res.choices[0].message.content)
import OpenAI from "openai";

const client = new OpenAI({ baseURL: "http://localhost:8000/v1", apiKey: "local" });
const res = await client.chat.completions.create({
  model: "zai-org/GLM-5.3-Flash",
  messages: [{ role: "user", content: "Write a Python function that parses a GGUF header." }],
});
console.log(res.choices[0].message.content);
Desktop
macOS
(M1 or better)
Download
Windows
(x64)
Download
Linux
(x86_64)
Download

Frequently asked questions

GLM-5.3-Flash is an open-weight Mixture-of-Experts model from Z.ai, published on 2026-08-25 under the MIT license. Z.ai describes it as the first natively multimodal model in the GLM-5 series, with 320B total parameters and 18B active per token; the repo metadata reports about 321B across all weights. The architecture is registered as glm5_next and the config carries both a text and a vision section, along with image and video token ids.

Z.ai states no hardware target, so the reference points are the published weight sizes. The default repository ships FP8 weights at 328.3 GB across 62 safetensors shards, and the separate BF16 repository is 642.7 GB across 120 shards. Those figures are weights only, before any KV cache, so this is multi-GPU server territory rather than a workstation model.

Yes. The weights are released under the MIT license, one of the most permissive licenses in open source. It allows commercial use, modification and redistribution with no royalties, as long as the license notice is kept. That is unusual for a model of this size, since large open-weight releases often ship under custom community licenses instead.

Z.ai combines sparse and linear attention in one hybrid stack, which it calls a first for the GLM series and credits with sharply reducing long-context serving costs while preserving precise long-context capabilities. The model also adopts Manifold-Constrained Hyper-Connections to improve scaling efficiency. Structurally it runs 45 layers with 288 routed experts, of which 8 are active per token, at a hidden size of 4096. Pre-training used a 30T-token multimodal corpus.

The config file sets max_position_embeddings to 1,048,576 tokens, so 1M. Worth noting that the model card itself states no context length anywhere; the figure comes from the configuration shipped with the weights. Z.ai does say the hybrid sparse and linear attention design exists to cut long-context serving costs while keeping long-context accuracy.