GLM-5.3-Flash: Specs, Pricing and MIT Open Weights

For twelve days in August 2026, the most-used model on OpenRouter had no name attached to it. "Ox Alpha" showed up unannounced, took text, images and video, ran a million-token context, and cost nothing. On 26 August, Z.ai claimed it: Ox Alpha was GLM-5.3-Flash, and the weights went up on Hugging Face under an MIT licence the same day.

This page is the verified specification. Every number below comes from a source we opened directly — the Hugging Face model card and its config.json, Z.ai's developer documentation and pricing page, the OpenRouter listing, and Artificial Analysis. Where sources disagree, we say so rather than picking the flattering one.

What is GLM-5.3-Flash?

GLM-5.3-Flash is a sparse mixture-of-experts language model with native vision. The headline confusion with every MoE release is the parameter count, so take it slowly: the model stores 320 billion parameters, but only routes about 18 billion of them per token. You need enough memory for all 320B; you pay the compute cost of 18B. That gap is the entire product.

The attention layout is the part worth understanding. Of the 45 layers, only every fourth one (layers 3, 7, 11 … 43) uses full sparse attention; the remaining 34 use linear attention with a short convolution and a top-k selection of 2,048 tokens. Linear attention scales with sequence length instead of its square, so the long-context bill collapses — Z.ai's launch material puts it at roughly 3× less attention compute and a 4.4× smaller KV cache than GLM-5.3, adding up to about a 3× end-to-end serving improvement. The sparse layers are what stop it losing the plot at long range, which is the failure mode pure-linear models have historically had.

One honest caveat on the context number. The architecture is configured for 1,048,576 positions, Z.ai's docs market it as "1M", and OpenRouter lists the served window as 1,310,720 tokens. But the model card's own long-context evaluations run to 300,000 tokens. Treat 1M as the ceiling the model will accept, not as a depth Z.ai has published quality numbers for.

How does GLM-5.3-Flash differ from GLM-5.3 and GLM-5.2?

These are three genuinely different models, and the licensing is not uniform across them — which matters if you are planning to self-host or resell.

Two things in that table are load-bearing. First, Flash is not a distilled GLM-5.3. It is a shallower, wider-expert, natively multimodal architecture — 45 layers against 78, 288 experts against 256, a different attention scheme entirely. It shares a family name, not a lineage.

Second, the licence split. GLM-5.3's full weights were held back at launch for a two-week safety evaluation, because the model turned out to be unusually strong at offensive-security reasoning — not just spotting isolated bugs but planning exploitation chains. When those weights did land on 28 August, they came under a bespoke glm-5.3 licence rather than MIT. Its notable clause requires any model-as-a-service operator with more than US$10 billion of revenue over any twelve consecutive months to pass a Z.ai security review before commercial hosting. Individual developers and normal companies are unaffected, but it is not the plain MIT grant GLM-5.2 shipped with.

GLM-5.3-Flash escaped all of that. It is straight MIT — the shortest, most permissive licence in common use, with no revenue triggers, no acceptable-use appendix and no field-of-use restriction. If your reason for caring about open weights is legal certainty rather than benchmark scores, Flash is the more interesting release of the two.

For the full picture on the previous generation, see our GLM-5.2 complete guide and the GLM-5.3 launch guide.

What does GLM-5.3-Flash cost?

Z.ai's published list price is $0.15 per million input tokens, $0.03 per million cached input tokens, and $0.50 per million output tokens. A launch promotion halves that — $0.075 / $0.015 / $0.25 — running until 24:00 on 9 September 2026 (UTC+8). Cached-context storage is free for the promotional period. There is no context-length tiering: the million-token window costs the same per token as the first thousand.

OpenRouter's listing currently shows $0.07125 input, $0.01425 cached and $0.2375 output, which is the promotional rate arriving through their margin. Budget against the $0.15 / $0.50 list price, not the promo, because the promo expires in days.

Here is how that sits against the rest of the cheap tier, all list prices per million tokens:

DeepSeek V4-Flash is cheaper on raw tokens. But raw token price is the wrong axis, because output volume varies enormously between models and reasoning models emit a lot of tokens. Artificial Analysis's cost-per-task figure is the more honest number, and it puts GLM-5.3-Flash at roughly $0.09 to complete one task in its Intelligence Index — while scoring 57 on that index. That is the story: not "cheapest", but the cheapest thing at this capability level, and roughly one tenth of what GLM-5.2 costs to run for meaningfully better results.

How good is GLM-5.3-Flash?

Split the evidence into what Z.ai measured and what someone else measured. Z.ai's own reported scores, from the Hugging Face model card and launch materials:

Vendor benchmarks are marketing until someone reproduces them. The independent read comes from Artificial Analysis, which scores GLM-5.3-Flash at 57 on its Intelligence Index — fourth of 111 models in its comparison set, against a median of 29 for open-weight models of similar size. Measured throughput is 43.8 output tokens per second with a 1.54-second time to first token, which ranks it 51st on speed. It is not a fast model in wall-clock terms; it is a cheap model that happens to be smart.

One number you will see quoted that you should not trust: the viral "80% on DeepSWE" claim from Ox Alpha's stealth period. That came from a developer running a 10-task subset and getting 8 passes. Ten tasks is not a benchmark — the confidence interval on 8/10 spans most of the plausible range, and the models it was compared against were evaluated on different samples. Z.ai's own audited DeepSWE v1.1 figure is 63.4, and that is the number to use.

Is GLM-5.3-Flash really multimodal?

Yes, and the distinction from bolted-on vision is real. Most "multimodal" open models are a text model with a vision encoder grafted on after the fact, aligned through a projection layer trained on a comparatively small image-text set. GLM-5.3-Flash was pre-trained on a 30-trillion-token multimodal corpus, so image and video tokens were in the mixture from the start rather than being translated into it afterwards.

Practically, the Z.ai API accepts video, image, text and file input and returns text. The vision tower is a 24-block encoder at 1,024 hidden width, processing 448×448 images in 14×14 patches, with dedicated image and video special tokens reserved in the vocabulary (IDs 154830–154855). Z.ai's release notes highlight interface and GUI understanding specifically — screenshots, dashboards, document layouts — which is the capability that makes it useful for computer-use agents rather than just image captioning.

It is the first natively multimodal model in the GLM-5 series. GLM-5.2 and GLM-5.3 are both text-only, which is exactly why Ox Alpha's video support was the clue that broke the fingerprinting: the behaviour matched Z.ai's stack but no released Z.ai model could take video.

How do you use GLM-5.3-Flash?

Two hosted routes, both OpenAI-compatible. Direct from Z.ai, the endpoint is https://api.z.ai/api/paas/v4/chat/completions and the model string is glm-5.3-flash:

curl https://api.z.ai/api/paas/v4/chat/completions \
  -H "Authorization: Bearer $ZAI_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "glm-5.3-flash",
    "messages": [{"role": "user", "content": "Summarise this repo'\''s test strategy."}],
    "temperature": 1,
    "top_p": 0.95,
    "reasoning_effort": "max",
    "stream": false
  }'

Through OpenRouter the model ID is z-ai/glm-5.3-flash, which gives you provider failover and one billing relationship across models:

from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key="YOUR_OPENROUTER_KEY",
)

resp = client.chat.completions.create(
    model="z-ai/glm-5.3-flash",
    messages=[{"role": "user", "content": "Explain this stack trace."}],
    temperature=1,
    top_p=0.95,
)
print(resp.choices[0].message.content)

Three settings matter. Z.ai recommends temperature: 1 and top_p: 0.95 — not the 0.2-ish temperature people habitually use for code, and departing from it degrades output. reasoning_effort takes low, high or max, defaulting to max; drop it to low for latency-sensitive work, since output tokens are what you are actually paying for. And if you are building a chat product, pass clear_thinking=true explicitly — it defaults to false, and you will otherwise ship reasoning traces to your users. Thinking mode cannot be turned off entirely. Function calling and JSON-formatted responses are supported; strict JSON-schema validation is not.

Can you run GLM-5.3-Flash on your own hardware?

You can, with caveats, and it is a bigger topic than fits here — we cover it properly in how to run GLM-5.3-Flash locally. The short version: the released FP8 weights are around 306 GiB before KV cache, so full-precision serving is a multi-GPU server job via SGLang, vLLM or TokenSpeed, all of which have official recipes on the model card. For single-machine use, Unsloth publishes dynamic GGUF quantisations that fit a 128 GB machine at 1-bit and 2-bit, with 4-bit wanting roughly 192–256 GB.

The blocker as of now is tooling, not hardware. The glm5_next architecture has not landed in mainline llama.cpp, so the GGUFs need Unsloth's branch. Everything downstream — LM Studio, Ollama's local runner — is waiting on that merge. Ollama currently offers only a glm-5.3-flash:cloud tag, which is a hosted passthrough, not local inference. Expect that to change quickly; check before assuming it hasn't.

Was Ox Alpha GLM-5.3-Flash?

Yes. Ox Alpha appeared on OpenRouter and OpenCode on 20 August 2026 with no owner attached, free to use, with a million-token context and image and video input — and promptly climbed the usage charts. Z.ai confirmed the identity when it launched GLM-5.3-Flash on 26 August, and noted the stealth period had been served entirely on domestically produced Chinese AI chips.

The community fingerprinting was largely right before the reveal: tokenizer behaviour, video encoder quirks and API configuration all pointed at Z.ai's stack. The one thing that confused everybody was the multimodality, since GLM-5.3 is text-only — which is precisely why the answer was an unreleased sibling rather than a rebadge. Our Ox Alpha stealth model guide has the full timeline, including how the fingerprinting was done.

Should you use GLM-5.3-Flash?

A simple decision rule. If you are running high-volume agentic or long-context work — code review, document extraction, GUI automation, anything that burns tokens by the million — GLM-5.3-Flash is currently the strongest thing available at its price, and MIT weights mean you are not locked to Z.ai's API if the pricing or the terms move. If you need the last few points of coding capability, GPT-5.6 Terra and Claude Opus 4.8 still edge it on Terminal-Bench, and you should pay for them on the work that justifies it. If you simply need the cheapest competent tokens and do not need vision, DeepSeek V4-Flash is less than half the price.

What has genuinely changed is the floor. A 320B natively multimodal model with a million-token context, MIT-licensed, at $0.15 per million input tokens is a different baseline than the one that existed a month ago. If you are choosing an LLM stack right now, our open-source LLM landscape guide maps the current field. And if you are building on top of models like this and need engineers who have already shipped agentic systems in production, Codersera places vetted remote developers who have.

FAQ

What is GLM-5.3-Flash?

GLM-5.3-Flash is Z.ai's first natively multimodal model in the GLM-5 series, released on 26 August 2026. It is a mixture-of-experts model with 320B total parameters and 18B active per token, built on a hybrid linear-plus-sparse attention architecture, with a 1,048,576-token context window and support for text, image, video and file input. Weights are on Hugging Face under the MIT licence.

Is GLM-5.3-Flash open source?

The weights are openly released under the MIT licence — the most permissive licence in common use, with no revenue thresholds or field-of-use restrictions. This differs from the full GLM-5.3 model, whose weights arrived on 28 August 2026 under a bespoke glm-5.3 licence requiring a Z.ai security review for model-as-a-service operators above US$10 billion in revenue.

How much does GLM-5.3-Flash cost?

Z.ai's list price is $0.15 per million input tokens, $0.03 per million cached input tokens, and $0.50 per million output tokens, with no context-length tiering. A launch promotion halves those rates to $0.075 / $0.015 / $0.25 through 24:00 on 9 September 2026 (UTC+8). OpenRouter currently reflects the promotional pricing at $0.07125 input and $0.2375 output.

How many parameters does GLM-5.3-Flash have?

320 billion total parameters, of which roughly 18 billion are active per token. It has 45 layers, 288 routed experts with 8 selected per token plus 1 always-active shared expert, and a hidden size of 4,096. You need memory for all 320B — about 306 GiB in the released FP8 format — but only pay compute for the 18B that fire.

Is GLM-5.3-Flash better than GLM-5.2?

On Z.ai's own reported benchmarks, substantially: DeepSWE v1.1 rises from 46.2 to 63.4 and AutomationBench from 26.2 to 48.8, at roughly one-tenth the API price. GLM-5.3-Flash also adds native image and video input, which GLM-5.2 lacks entirely. GLM-5.2 remains a much larger model, so for pure text tasks the gap is narrower than the headline numbers suggest.

Was Ox Alpha GLM-5.3-Flash?

Yes. Ox Alpha was the anonymous model that appeared on OpenRouter and OpenCode on 20 August 2026 and topped usage charts while free. Z.ai confirmed at the 26 August launch that Ox Alpha was GLM-5.3-Flash running its stealth evaluation period, served on domestically produced Chinese AI chips. Community fingerprinting had correctly identified Z.ai's stack beforehand.

Can I run GLM-5.3-Flash locally?

Yes, but it needs serious memory. FP8 weights are around 306 GiB, so full-precision serving needs a multi-GPU server via SGLang, vLLM or TokenSpeed. Unsloth's dynamic GGUF quantisations run on a 128 GB machine at 1-bit and 2-bit, with 4-bit wanting 192–256 GB. The glm5_next architecture is not yet in mainline llama.cpp, so Ollama and LM Studio cannot serve it locally yet.

Does GLM-5.3-Flash really have a 1M-token context?

Architecturally yes — max_position_embeddings is 1,048,576, Z.ai's docs market a 1M window, and OpenRouter lists 1,310,720 tokens served. But the model card's published long-context evaluations only run to 300,000 tokens, so 1M is the accepted input ceiling rather than a depth Z.ai has demonstrated quality at. Test your own workload before relying on it past 300K.