GLM-5.3-Flash

Name: GLM-5.3-Flash
Author: zai-org

by zai-org

Z.ai's first natively multimodal GLM-5 model: 320B total, 18B active, MIT licensed, built for agents that read screens and documents

320B total (MoE), 18B active per token; FP8 weightsparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

zai-org/GLM-5.3-Flash

Feature URI

Overview

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series from Z.ai, released under the MIT license with FP8 weights (repository created 2026-08-25, last updated 2026-09-04). It is a mixture-of-experts model with 320B total parameters and 18B active per token, trained on a 30T-token multimodal corpus. The model card describes a hybrid architecture that combines sparse and linear attention to cut long-context serving cost, plus Manifold-Constrained Hyper-Connections for scaling efficiency, and reports that it outperforms GLM-5.2 across benchmarks while approaching Claude Opus 4.8 on coding and agentic benchmarks at a tenth of the price. For a retrieval pipeline it belongs at the step where an agent reads a screenshot, chart or document page and acts on it, with the text it produces embedded for search by a separate embedding model.

Architecture

Mixture-of-experts transformer (Glm5NextForConditionalGeneration) with 320B total and 18B active parameters, a hybrid of sparse and linear attention, and Manifold-Constrained Hyper-Connections (mHC). Ships as FP8 safetensors in 62 shards with a chat template that exposes a reasoning_effort control (low, high, max) and a clear_thinking switch. Served through SGLang, vLLM, Transformers, KTransformers, TokenSpeed and Unsloth per the model card.

Mixpeek SDK Integration

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors), so GLM-5.3-Flash runs on your side and the
// text it produces is upserted through POST
// /v1/namespaces/{namespace_id}/documents/upsert. On Enterprise the other
// path is to upload the weights instead: POST /v1/namespaces/{id}/models
// accepts the huggingface format and a custom plugin loads them.
const res = await fetch(
  "https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer API_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      collection_id: "col_your_collection",
      documents: [
        {
          document_id: "page-00017",
          // The model produces text, so it lands in payload. Give the
          // collection a text vector index and embed that text to make it
          // searchable as well as filterable.
          payload: { figure_description: modelOutput, source_key: "reports/q3/page-17.png" },
          vectors: { "multimodal-embedding": embeddingOfModelOutput },
        },
      ],
    }),
  },
);

// Managed alternative, if this exact model is not the requirement:
// universal_extractor@v1 runs google/gemini-embedding-2
// (3072-d) over a bucket, with no inference of your own.

Capabilities

Image plus text input with text output, including screenshots, charts and document pages
Agentic and coding tasks: the card reports gains over GLM-5.2 and results approaching Claude Opus 4.8 on coding and agent benchmarks
Long-context serving at lower cost through the sparse-plus-linear attention hybrid
Adjustable thinking budget through reasoning_effort, with max as the default for benchmark reproduction
MIT license, FP8 weights, English and Chinese