GLM 5.3 Flash API

GLM 5.3 Flash is a natively multimodal large language model released and open-sourced by Zhipu (Z.ai) on August 26, 2026. It is the cost-efficient variant of the GLM-5.3 series and was stealth-tested as "Ox Alpha" on OpenRouter and OpenCode before launch, topping the usage charts. It accepts text, image, video, and file input with text output, offers a 1,048,576-token (1M) context window, and has 320B total / 18B active parameters. It scores 57 on the Artificial Analysis Intelligence Index (AA Index), level with Claude Opus 4.8, and its weights are open-sourced under the MIT license.

On iCreat, you can call glm-5.3-flash through an API compatible with OpenAI Chat Completions. The endpoint supports standard system and user messages, thinking mode, streaming output, and function calling, letting applications already built on the OpenAI SDK connect directly.

Model Positioning

GLM 5.3 Flash is the first natively multimodal model in the GLM-5 series, positioned to "keep strong performance while driving cost down": it is priced at roughly one-tenth of the flagship GLM-5.3 and targets coding agents, long-context workloads, and cost-sensitive high-throughput scenarios.

For teams that need one model to handle text, images, video, and files, it is currently the only option in Zhipu's GLM lineup — both GLM-5.3 and GLM-5.2 accept text input only. It offers thinking mode (controlled via thinking.type), function calling, and structured output. For the hardest tasks that push the intelligence ceiling (such as defensive cybersecurity), choose GLM-5.3.

Core Capabilities

Natively multimodal understanding

GLM 5.3 Flash was pre-trained from scratch on a 30-trillion-token multimodal corpus and natively reads text, image, video, and file input — the first GLM-5 series model to do so without a separate vision adapter. Its vision capability plugs into coding-agent loops: the model inspects the rendered output of web pages, GUIs, games, or 3D scenes and keeps editing code based on what it sees.

Ultra-long context

A 1M-token context window lets you load an entire codebase, massive logs, or multiple long documents in a single request for project-level refactoring plans, root-cause analysis, and clause-conflict detection. Maximum output is 131,072 tokens (128K).

Hybrid attention architecture

It combines sparse attention with linear attention, plus Manifold-Constrained Hyper-Connections (mHC) and an IndexPool mechanism. Versus GLM-5.3, attention compute drops about 3x and the KV cache shrinks about 4.4x, bringing long-context serving cost to roughly one-third.

Thinking mode and reasoning

Thinking mode is controlled via the thinking.type parameter (enabled/disabled), and reasoning_effort adjusts reasoning investment (e.g., low). Suited to programming, math, and agent tasks that need multi-step reasoning.

Function calling and agentic capabilities

It supports function calling and structured output for connecting the model to external tools and systems, and ranks in the top tier of open-source models on real-terminal and software-engineering benchmarks such as Terminal-Bench and DeepSWE.

OpenAI compatibility

Fully compatible with the OpenAI Chat Completions API — standard parameters such as temperature, max_tokens, and top_p work without adaptation. Connect directly through iCreat with the OpenAI Python/Node.js SDKs.

Pricing

Use Cases

• Coding agents: Plug into OpenCode, Cursor, and other coding tools for multi-file refactors, repo-wide bug hunts, and long-context code review.
• Vision-in-the-loop coding: The model views rendered web pages, GUIs, games, or 3D scenes and iterates on code — ideal for frontend and interactive application development.
• Long-context analysis: Load a full codebase, massive logs, or multiple long documents into the 1M window at once for root-cause analysis and conflict detection.
• Multimodal content understanding: Combine text, images, video, and files in one request for cross-modal analysis.
• High-throughput workloads: Input pricing at roughly one-tenth of the flagship suits large-scale batch processing and cost-sensitive production environments.
• Self-hosted deployment: MIT open weights downloadable from Hugging Face can be served locally with SGLang, vLLM, and other inference frameworks with no licensing restrictions.

Model Comparison

GLM 5.3 Flash vs. GLM 5.3

GLM 5.3 Flash vs. GLM 5.2 and GLM 4.7 Flash

Why Choose GLM 5.3 Flash?

When your workload is dominated by coding agents and long-context processing, and repeated context drives most of the cost, GLM 5.3 Flash is one of the best value frontier-tier options available. The $0.03 per million tokens cached-read rate makes re-sent context nearly free, input/output prices are roughly one-tenth of the flagship GLM-5.3, and its AA Index score of 57 still matches Claude Opus 4.8.

Versus GLM-5.2, it comprehensively beats the larger previous flagship on software-engineering benchmarks (DeepSWE v1.1: 63.4 vs 46.2) while adding native multimodal input at about one-tenth the price. Versus GLM-5.3, it delivers near-flagship intelligence at roughly one-tenth the cost plus image, video, and file input, trailing only on the hardest tasks (AA Index 60 for the flagship).

Through iCreat, teams connect via the OpenAI-compatible endpoint without deploying Zhipu's native SDK. The MIT open weights also let teams with GPU resources self-host and eliminate per-token costs entirely.

Specifications

Architecture

GLM 5.3 Flash uses a Mixture-of-Experts (MoE) architecture with 320B total parameters, 18B active per token, across 45 layers. It is the first open-source frontier model to adopt a hybrid sparse-plus-linear attention design, paired with Manifold-Constrained Hyper-Connections (mHC) for scaling and an IndexPool mechanism that compresses four indexer key vectors into one weighted pool to cut long-context latency.

Versus the previous-generation GLM-4.5, total parameters dropped from 355B to 320B, active parameters from 32B to 18B, and layers from 92 to 45; versus the full GLM-5.3, attention compute falls about 3.01x and the KV cache shrinks about 4.44x. The model was pre-trained on 30 trillion multimodal tokens. Serving is built on SGLang with an EPD (encoding/prefill/decode separation) architecture, lifting service performance roughly 3x over the initial baseline, with all traffic carried by domestic Chinese chip clusters.

Notes

When called through iCreat's OpenAI-compatible endpoint, the model outputs text only — no speech or image output.

The 1M-token context is a shared request space. Image and video inputs consume many tokens; reserve room for system instructions, conversation state, and the final answer, and assess the token budget before sending long videos.

With thinking mode enabled, responses consume reasoning tokens. For latency- and cost-sensitive high-throughput scenarios, control reasoning investment via reasoning_effort: "low"; raise the level for complex tasks.

Agent loops that repeatedly re-send large context blocks should leverage caching: the cached-read price of $0.03 per million tokens significantly reduces the cumulative cost of multi-turn agents.

Multimodal input requires accessible URLs; local files cannot be uploaded directly. Video URLs must point to actual video files, and image URLs to actual image files.

The AA Index score of 57 trails the flagship GLM-5.3's 60. For the hardest tasks such as defensive cybersecurity, evaluate GLM-5.3 first; when evaluating GLM 5.3 Flash, observe full task completion and measure code-fix rates, agent task success rates, multimodal understanding accuracy, latency, and token usage.

FAQ

What is the difference between GLM 5.3 Flash and GLM 5.3?

GLM 5.3 Flash is a natively multimodal model (text, image, video, and file input) priced at roughly one-tenth of GLM-5.3, with an AA Index score of 57; GLM-5.3 is a text-only flagship with an AA Index score of 60, strong at complex coding, defensive cybersecurity, and long-horizon tasks. Both offer a 1M context and MIT open weights. Choose Flash for high-throughput and multimodal scenarios; choose GLM-5.3 for the hardest tasks.

What is "Ox Alpha"?

Ox Alpha was the anonymous testing codename for GLM 5.3 Flash before its release. Zhipu offered it for free on OpenRouter and OpenCode under that name; it topped OpenRouter's usage chart on day one and processed 42 trillion tokens in six days, with all requests served by domestic Chinese chip clusters.

Can I call it directly with the OpenAI SDK?

Yes. Set base_url to https://api.icreat.ai/llm/openai/v1 and api_key to your iCreat API Key — no adaptation needed. Standard parameters such as temperature, max_tokens, and top_p all work.

How do I enable streaming output?

Add "stream": true to the request body. Responses return chunk by chunk in SSE (Server-Sent Events) format, each chunk containing delta.content, with the final chunk's finish_reason set to stop.

How do I control thinking mode?

Use the thinking.type parameter: enabled to turn it on, disabled to turn it off. When enabled, reasoning_effort adjusts reasoning investment (e.g., low). Lower or disable thinking for simple tasks to save latency and cost; raise the level for complex tasks.

How does it compare with GLM 5.2?

GLM 5.2 is the previous flagship released on June 16, 2026 (744B MoE, 1M context, text-only). GLM 5.3 Flash comprehensively surpasses GLM-5.2 in benchmarks and real-world use despite its smaller size (DeepSWE v1.1: 63.4 vs 46.2), costs about one-tenth as much, and adds native multimodal input.

Can I self-host it?

Yes. The weights are open-sourced on Hugging Face under the MIT license and can be deployed with inference frameworks such as SGLang and vLLM, including commercially. The hybrid attention architecture shrinks the KV cache by about 4.4x, making 1M-context self-hosting more feasible in terms of GPU memory.