GLM-5.3-Flash: Z.ai's Multimodal Model at One-Tenth the Price

Z.ai launched GLM-5.3-Flash on August 26, 2026 — the first natively multimodal model in its GLM-5 series, released under the MIT license with a 1M-token context window (Z.ai on X). It's a 320B-A18B model that Z.ai says beats GLM-5.2 across coding and agent tests at roughly a tenth of the price. Here's what's actually new, the benchmarks that matter for agents, what it costs, and how it differs from GLM-5.3.

What is GLM-5.3-Flash?

GLM-5.3-Flash is a Mixture-of-Experts model with 320 billion total parameters and 18 billion active parameters per token (TestingCatalog). Z.ai positions it as the lower-cost, higher-efficiency tier of the GLM-5 family: strong coding and agentic performance, but built to run cheaply and respond fast.

Two things set it apart from earlier GLM releases. It's the first natively multimodal model in the line — vision is baked into the training, not bolted on — and it ships under the MIT license with open weights on Hugging Face. If the name "Ox Alpha" sounds familiar, that's because this is the same model: it ran anonymously as ox-alpha on OpenCode and OpenRouter for about a week before the reveal.

The hybrid architecture: cheaper attention at 1M tokens

The efficiency story starts with the attention design. GLM-5.3-Flash is, per Z.ai, the first open-source frontier model to combine sparse attention and linear attention in one hybrid architecture — linear attention for local dependencies, sparse attention for the relevant global context (Z.ai docs).

The payoff is concrete. Compared with GLM-5.3, Z.ai reports roughly 3x less attention compute and a 4.4x smaller KV cache (TestingCatalog). At context lengths reaching one million tokens, a component Z.ai calls IndexPool compresses groups of indexer key vectors to hold down latency and memory. That's what makes a 1M-token window usable in practice rather than just on a spec sheet — long-horizon agents can keep more context loaded without the cost blowing up.

The model was trained on a 30-trillion-token multimodal corpus, and Z.ai says it starts from a newly trained base rather than reusing GLM-5.2's — a departure from the GLM-5.3 recipe, which reused the GLM-5.2 base and scaled only post-training.

GLM-5.3-Flash benchmarks

All numbers below are vendor-reported or from third-party trackers; harnesses and settings differ, so read them as directional rather than head-to-head.

On coding and agent suites versus GLM-5.2, the gains are large rather than incremental (OfficeChai):

• DeepSWE v1.1: 46.2 → 63.4
• AutomationBench: 26.2 → 48.8 (nearly double)
• Terminal-Bench 2.1: 84.3 — ahead of DeepSeek-V4-Vision-Exp and within striking distance of Claude Opus 4.8's 85.0

On the same Terminal-Bench test, GPT-5.6 Terra (87.4) and Gemini 3.7 Flash (85.8) still edge ahead, but the margin has narrowed sharply from where GLM-5.2 sat. On Z.ai's own internal Code Bench v1.0 — run inside Claude Code — GLM-5.3-Flash tops out at 29.0 on max effort against Opus 4.8's 29.5, which the company points to as evidence its flash-tier model now trades blows with a frontier model on coding.

On the independent Artificial Analysis Intelligence Index, GLM-5.3-Flash scores 57 — well above the median for reasoning models in its price tier. The honest caveat: internal benchmarks like Code Bench can't be independently replicated, and the model doesn't claim to top GPT-5.6 Sol or Fable 5 on the hardest public suites. It's a value play, not a leaderboard-topping play.

Native multimodal: vision inside the coding loop

The multimodal capability isn't a separate feature — it's wired into how the model works. Z.ai trained GLM-5.3-Flash to inspect rendered interfaces, gameplay, and 3D output, then assess and revise its own work from that visual feedback (TestingCatalog).

That "look at the result, then fix it" loop matters for agents that build UIs or dashboards: the model can see the rendered page, notice it's wrong, and iterate. The same approach extends beyond code to documents, spreadsheets, presentations, and meeting materials — so the model can reason across text, images, and structure, and deliver finished PPTX, PDF, DOCX, and XLSX files end to end (Z.ai docs).

How much does GLM-5.3-Flash cost?

Pricing is the headline. Z.ai's standard API rates are $0.15 per 1M input tokens and $0.50 per 1M output, with cached input at $0.03 (Z.ai on X). Third-party providers list it even lower — OpenRouter shows $0.075 input / $0.25 output (OpenRouter).

For GLM Coding Plan subscribers, it's already live with 3x the usable quota of GLM-5.3 (TestingCatalog). That's the practical difference from the pricier tiers: you can run long agent loops without watching a token meter.

GLM-5.3-Flash vs GLM-5.3

They're built for different jobs. GLM-5.3 is the reasoning-heavy coding model — the one that grew an unplanned cybersecurity skill and had its weights staged for safety review. GLM-5.3-Flash is the efficient, multimodal, MIT-licensed variant meant to run cheaply at scale.

The quick contrast:

• Modality: GLM-5.3 is text-focused; GLM-5.3-Flash is natively multimodal.
• Architecture: GLM-5.3 reuses the GLM-5.2 base; GLM-5.3-Flash uses a newly trained base with hybrid sparse + linear attention.
• Efficiency: GLM-5.3-Flash reports ~3x less attention compute and a 4.4x smaller KV cache than GLM-5.3.
• License and weights: GLM-5.3-Flash ships MIT-licensed with open weights on day one; GLM-5.3's weights were staged.
• Cost: GLM-5.3-Flash is priced for volume — roughly a tenth of GLM-5.2's cost, per Z.ai.

If you need the deepest reasoning on a hard problem, GLM-5.3 is the pick. If you're running multimodal or long-horizon agents where cost-per-task decides whether the workflow is viable, GLM-5.3-Flash is the one to test.

What it means if you build AI agents

A few practical takeaways:

• This is a cost lever for all-day agents. At $0.075–$0.15 per 1M input tokens with a real 1M-token window, GLM-5.3-Flash makes long-running agent loops affordable in a way frontier models aren't.
• Multimodal changes what agents can verify. A model that can see a rendered interface and fix it closes a loop that text-only agents can't — useful for UI, dashboard, and document work.
• MIT license removes friction. Open weights under a permissive license mean teams with data-residency or vendor-review rules can self-host from day one; local deployment supports SGLang, vLLM, and TokenSpeed.
• Treat the internal benchmarks as a starting point. Code Bench numbers are private and fresh. Independent testing is only just beginning now that the weights are out.

Run a model like GLM-5.3-Flash in your own AI workforce

GLM-5.3-Flash is built for exactly the work Eigent orchestrates: cheap, long-horizon, multimodal agent loops that run end to end instead of answering one prompt. Eigent is an open-source, local "Cowork" desktop app that turns models like these into a multi-agent workforce — from reviewing GitHub PRs to running a self-hosted AI coding agent you fully own. Bring your own model, keep the work on your machine. Download Eigent and build your own agent workflows today.