GLM-5.3-Flash: Smart 320B Open Weights, Surprising Price

GLM-5.3-Flash is the model that spent a week at the top of OpenRouter under a codename nobody could crack. On 26 August 2026, Zhipu AI — trading as Z.ai — confirmed that the free, anonymous “Ox Alpha” endpoint developers had been hammering all week was in fact GLM-5.3-Flash, and published the weights on Hugging Face under an MIT licence the same day.

That is an unusual sequence. Most laboratories announce a model, publish a benchmark table, and hope somebody uses it. Z.ai inverted the order: it shipped an unbranded endpoint, let the developer community stress it for six days without knowing whose large language model they were testing, and only then attached a name to the results.

The numbers that came out of that week are why this launch matters. GLM-5.3-Flash is a 320-billion-parameter mixture-of-experts model that activates only 18 billion parameters per token. It scores 57 on the Artificial Analysis Intelligence Index — the same figure as Claude Opus 4.8 — while its blended API price sits near one-tenth of Z.ai’s own flagship and a small fraction of the frontier labs’ rates.

Founder Jie Tang added the detail that turned a model release into a geopolitical story. Writing on 27 August, he summarised the launch as: “Ox Alpha = GLM-5.3 Flash AA = 57, 1/100 frontier price, Powered by pure Chinese chips. Delivered nearly 20% weekly token share (no. 1) on OpenRouter.”

This article works through what GLM-5.3-Flash actually is, what the architecture does differently, what the benchmarks support and what they do not, what it really costs, how credible the domestic-silicon claim is, and what any of it changes for a business choosing a model this quarter.

Some of what follows is less flattering than the launch coverage. The model is genuinely slow per token, the headline price is a promotion with an expiry date, and the hardware floor for self-hosting is high enough that “open weights” does not mean “runs on your server”. Those caveats matter more than the leaderboard position.

Table of contents

What GLM-5.3-Flash Actually Is
The Ox Alpha Run That Came Before the Name
Inside the GLM-5.3-Flash Architecture
What the GLM-5.3-Flash Benchmarks Show
What GLM-5.3-Flash Costs
The Chinese Chip Claim, Examined
Running GLM-5.3-Flash Yourself
What GLM-5.3-Flash Changes for Businesses
GLM-5.3-Flash: Frequently Asked Questions
References and Further Reading

What GLM-5.3-Flash Actually Is

Strip away the launch theatre and GLM-5.3-Flash is a sparse mixture-of-experts model built for cheap, long-running agentic work rather than for topping a reasoning leaderboard.

A 320B model that only ever uses 18B

The headline parameter count is 320 billion. The number that governs cost is 18 billion — the parameters actually activated for any given token. That ratio is what lets Z.ai serve GLM-5.3-Flash at a fraction of the price of a comparably capable dense model, and it is a sharper ratio than the company’s previous generation used.

The stack is 45 layers deep: three dense feed-forward layers at the base, then 42 mixture-of-experts layers. Each of those routes across 288 experts plus one shared expert, activating eight routed experts per token.

Natively multimodal, not bolted on

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series. Text, images, video, visual documents and interleaved combinations of all four are handled by one model trained on a 30-trillion-token multimodal corpus, with a 24-layer vision transformer trained jointly rather than in a separate pre-training phase.

That distinction is not cosmetic. Vision encoders grafted onto a finished text model tend to behave like a translation layer; a jointly trained encoder lets the model reason across a chart, a spreadsheet and a screenshot in one pass. Z.ai leans on this for its Browser Use and Computer Use features.

The licence is the part that travels

The weights are on Hugging Face as zai-org/GLM-5.3-Flash under an MIT licence — the permissive end of the spectrum, with no field-of-use restrictions and no acceptable-use rider. Two checkpoints ship: a native FP8 build of roughly 331 GB and a BF16 build near 640 GB.

An MIT licence on a model at this capability level is a meaningful event regardless of what the benchmark table says. It means the weights can be fine-tuned, redistributed and embedded commercially without a negotiation, which is not true of most models scoring anywhere near it.

Specification	GLM-5.3-Flash
Total parameters	320 billion
Active parameters per token	18 billion
Layers	3 dense + 42 mixture-of-experts
Experts	288 routed + 1 shared, 8 routed active
Context window	1,048,576 tokens
Maximum output	131,072 tokens
Pre-training corpus	30 trillion multimodal tokens
Licence	MIT
Checkpoints	FP8 (~331 GB), BF16 (~640 GB)

The Ox Alpha Run That Came Before the Name

The week before the announcement is the most interesting part of the story, and it is the part that will be copied.

Six days at the top of a public leaderboard

From around 20 August, an unbranded endpoint called Ox Alpha appeared free on OpenRouter and OpenCode, advertising a one-million-token multimodal context window and a 131,072-token output ceiling. Z.ai says it stood up capacity for roughly 100 trillion tokens a day.

Developers used it. By the time the model was named, it had processed 62 trillion tokens on OpenRouter, including 11 trillion in its first three days — the platform’s largest launch to date. It ranked first among coding systems at 10.3 trillion tokens, about 31% of weekly volume, and took close to 20 percentage points of OpenRouter’s total weekly token share on its own.

What an anonymous launch actually tests

Running a model unbranded removes the two things that usually distort early reception: the vendor’s benchmark table and the audience’s priors about the vendor. Nobody could dismiss Ox Alpha as a Chinese model or hype it as one, because nobody knew.

It also tests serving capacity honestly. A free endpoint under real load for six days is a harder infrastructure exercise than a launch-day announcement, and it produced usage data no internal evaluation could.

Z.ai had already trailed the reveal

We covered the unmasking as it happened in our piece on Z.ai as the lab behind Ox Alpha, and the earlier stage — when the mystery model was simply being tested for free — before either had a name attached. GLM-5.3-Flash is the answer to both.

The market reaction was immediate. Zhipu’s Hong Kong-listed shares closed more than 12% higher at HK$1,160 on Thursday 28 August, the session after the reveal.

Inside the GLM-5.3-Flash Architecture

The efficiency claims around GLM-5.3-Flash are not marketing gloss on a conventional design. The attention stack is where the work went.

Hybrid linear and sparse attention

Of the 45 layers, 34 use a linear attention variant that scales with sequence length rather than its square, and 11 use a sparse multi-head latent attention design with no positional encoding. The linear layers carry local dependencies cheaply; the sparse layers preserve global precision where it matters.

This is the trade every long-context model has to make. Pure quadratic attention gives the best fidelity and an impossible bill at a million tokens. Pure linear attention is affordable and loses long-range precision. The hybrid split is an engineering answer to that, not a theoretical one.

IndexPool and the KV cache problem

The specific innovation Z.ai names is IndexPool, which compresses indexer key vectors through weighted pooling. The reported effect against the full GLM-5.3 model is 3.01 times less attention compute and a 4.44 times smaller KV cache.

KV cache size is the practical constraint on long-context serving. It grows with context length and concurrency, and it is usually what makes a million-token window a menu item rather than something anyone actually uses. Cutting it by more than four times is how GLM-5.3-Flash makes that window ordinary.

FP8 as the shipping format, not an afterthought

The primary checkpoint is native FP8 with dynamic activation scaling, with BF16 offered as the alternative. Shipping FP8 first is a deliberate signal: the model is meant to be served, and the eight-bit numbers are the ones that were trained and evaluated.

The stack also includes manifold-constrained hyper-connections for residual-stream stability and a single multi-token-prediction draft layer that supports speculative decoding across five tokens.

The gap between servable and evaluated context

One honest caveat sits in the technical notes. The maximum servable context is 1,048,576 tokens, but evaluation was run at roughly 300,000 text tokens and 164,000 vision tokens. The million-token figure is a capacity, not a validated working range, and that distinction is worth carrying into any design that depends on it.

What the GLM-5.3-Flash Benchmarks Show

The published comparisons put GLM-5.3-Flash against its own predecessor, against Claude Opus 4.8, and against GPT-5.6 Terra.

Agentic and terminal work

On Terminal-Bench 2.1 the model scores 84.3, against 81.0 for GLM-5.2, 85.0 for Claude Opus 4.8 and 87.4 for GPT-5.6 Terra. That is within a point of Opus on a benchmark built around multi-step command-line tasks.

Terminal-Bench 2.1 score (bar width equals the score out of 100)

GPT-5.6 Terra 87.4

Claude Opus 4.8 85.0

GLM-5.3-Flash 84.3

GLM-5.2 81.0

The jump over its own predecessor

The more striking comparison is generational. On DeepSWE v1.1, GLM-5.3-Flash scores 63.4 against GLM-5.2’s 46.2 — a 17.2-point gain. On AutomationBench the gap is wider still: 48.8 against 26.2, which is 22.6 points.

DeepSWE v1.1 score (bar width equals the score out of 100)

GPT-5.6 Terra 69.6

GLM-5.3-Flash 63.4

Claude Opus 4.8 58.0

GLM-5.2 46.2

The composite index and its company

The Artificial Analysis Intelligence Index v4.1.1 places GLM-5.3-Flash at 57, fourth of 111 models in its comparison set, against a median of 29 for open-weight models of similar size. The index is a composite of nine evaluations including GDPval-AA v2, Terminal-Bench v2.1, SciCode, Humanity’s Last Exam, GPQA Diamond and AA-Omniscience.

The full evaluation cost $138.02 to run and produced 150 million output tokens, above the 110-million median. That verbosity is worth noting: a model that thinks in more tokens is cheaper per token but not automatically cheaper per task.

Where it does not win

Three places. It trails GPT-5.6 Terra on every headline coding benchmark in the table. It trails Claude Opus 4.8 on Humanity’s Last Exam with tools, and on Z.ai’s own internal Code Bench it reaches 29.0 at maximum effort against 29.5 for Opus. And its measured output speed is 49.4 tokens per second against a 64.8 median — genuinely slow.

Benchmark	GLM-5.3-Flash	GLM-5.2	Claude Opus 4.8	GPT-5.6 Terra
Terminal-Bench 2.1	84.3	81.0	85.0	87.4
DeepSWE v1.1	63.4	46.2	58.0	69.6
Agents’ Last Exam	26.3	20.4	27.0	28.0
AutomationBench	48.8	26.2	41.0	37.2
GDPval-AA v2 (Elo)	1773	1504	1582	1571

The multimodal side of the table

On the vision evaluations the pattern holds: 89.4 on CharXiv-R, 80.5 on MMVU, 78.4 on Toolathlon and 78.0 on Chartography. These are document, chart and video reasoning tasks rather than image captioning, which is the right emphasis for a model aimed at agent work.

GLM-5.3-Flash multimodal scores (bar width equals the score out of 100)

CharXiv-R 89.4

MMVU 80.5

Toolathlon 78.4

Chartography 78.0

What GLM-5.3-Flash Costs

Price is the argument Z.ai is actually making, and it needs reading carefully.

List price and the promotion

The list price is $0.15 per million input tokens, $0.03 per million cached input tokens and $0.50 per million output tokens. A 50% launch promotion halves the first and last of those through 9 September 2026, which is where the widely quoted $0.075 and $0.25 figures come from.

That expiry date matters. Any cost model built on promotional pricing needs a second column with the list rate in it, because the list rate is what applies from 10 September.

Against Z.ai’s own flagship

GLM-5.3 — the full model, not the Flash variant — lists at $1.40 per million input tokens and $4.40 per million output. Blended, GLM-5.3-Flash lands near $0.10 per million against roughly $0.90 for GLM-5.3, which is where the “one-tenth the price” line comes from.

Rate	GLM-5.3-Flash list	GLM-5.3-Flash promo	GLM-5.3
Input, per 1M tokens	$0.15	$0.075	$1.40
Cached input, per 1M	$0.03	$0.015	—
Output, per 1M tokens	$0.50	$0.25	$4.40
Blended, per 1M	~$0.10	—	~$0.90

The provider spread is larger than the discount

Eighteen providers serve GLM-5.3-Flash on OpenRouter and eleven are tracked by Artificial Analysis, and they do not perform alike. Output speed ranges from 47 to 273 tokens per second — a 484% spread between fastest and slowest.

Databricks measured 272.9 tokens per second with an 8.00-second time to first token, FriendliAI 242.1 at 9.51 seconds, Baseten 194.7 at 10.95 seconds. Blended prices across providers run from $0.05 to $0.12 per million tokens. Picking the provider is a bigger decision than picking the discount.

What “one-hundredth of frontier” means

Jie Tang’s claim is a ratio against the most expensive frontier endpoints, not against every commercial model. Read as a directional statement about where open-weight pricing has landed, it is fair. Read as a per-task cost comparison, it is not — a slower, more verbose model recovers some of that gap in tokens spent, and any serious cost optimization exercise has to measure completed tasks rather than token rates.

The Chinese Chip Claim, Examined

This is the part of the GLM-5.3-Flash story that moved a share price, and it deserves more precision than it usually gets.

What Z.ai actually said

The company says the Ox Alpha week was served entirely on domestically produced AI chips — a cluster reported at around 100,000 units. It declined to name the vendor, noting only that the silicon has less compute and less memory than the Nvidia equivalents.

Zhipu has previously stated that the GLM-5 family was trained on Huawei Ascend hardware without Nvidia parts, and it works with Huawei, Cambricon and Moore Threads. Cambricon said it had achieved day-zero compatibility for serving GLM-5.3-Flash.

What it does and does not prove

It demonstrates that a competitive frontier-adjacent model can be trained and served at scale on non-Nvidia, Chinese-manufactured accelerators. For anyone modelling export controls as a hard ceiling on Chinese capability, that is a real data point.

It does not prove parity. Serving a model is a lighter workload than training one, “less compute and memory” is the company’s own characterisation, and an unnamed chip cannot be independently verified. The 49.4 tokens per second measured output speed is consistent with hardware that is adequate rather than abundant.

Why the reveal moved a share price

Markets priced the combination, not the model. A competitive open-weight release, a leaderboard position, and a demonstration of domestic silicon independence arrived in the same announcement, and the shares closed up more than 12%.

We looked at the wider version of this question — who gets access to frontier systems and on whose hardware — in our piece on frontier AI access as a national security question.

Running GLM-5.3-Flash Yourself

Open weights and self-hostable are not the same claim, and the gap is wide here.

The hardware floor is high

Serving GLM-5.3-Flash needs Hopper-generation GPUs or newer. A100s and older are ruled out by the FlashInfer dependency. The realistic minimum is a single GB200 tray at tensor-parallel 4, around 386 GB of VRAM for the FP8 checkpoint, or an eight-way H100 or H200 node. BF16 at tensor-parallel 8 wants roughly 772 GB.

That puts self-hosting firmly in the datacentre, not the workstation. The MIT licence is valuable for fine-tuning, sovereignty and redistribution — it is not an invitation to run this on a spare server.

Serving frameworks

Support landed quickly: vLLM 0.27.0 or later with FlashInfer 0.6.17 or later, SGLang, TokenSpeed, and KTransformers for hybrid CPU-GPU serving. That breadth on day one is itself a signal of how much preparation went into the launch.

When the API is the better answer

For most teams it will be. At $0.15 per million input tokens, the arithmetic favouring self-hosting only works at very high sustained volume, and it stops working entirely once you price the engineering time. Self-host when you need the weights on your own infrastructure for data or regulatory reasons — not to save money.

Fine-tuning is the real MIT dividend

Where the licence earns its keep is adaptation. A permissively licensed 18-billion-active model with a strong agentic baseline is a good starting point for domain work, and that sits squarely in ordinary ml model development practice rather than in anything exotic. Post-training methods including supervised fine-tuning and reinforcement learning apply here as they would to any open checkpoint.

What GLM-5.3-Flash Changes for Businesses

Leaderboards are not procurement decisions. Here is what actually shifts.

Long-context work gets affordable

A million-token window at $0.15 per million input tokens changes which problems are worth attempting. Whole-repository code review, full contract sets, complete support-ticket histories — workloads previously priced out are now arithmetic that closes. That is the concrete change GLM-5.3-Flash introduces.

Cheap agents are still agents

Long-horizon agentic work is the stated design target, and the AutomationBench and Terminal-Bench numbers support it. But cheaper autonomous agents are not safer autonomous agents, and a model that costs a tenth as much to run will be given ten times as many chances to act. Scope, permissions and audit trails matter more, not less. Any sensible ai strategy treats a price drop as a governance event rather than only a budget one.

Data residency is a live question

Using the Z.ai API means sending data to a Chinese provider. For a lot of UK workloads that is a straightforward data protection blocker regardless of what the benchmarks say. The MIT weights give you an answer — self-host, or use a Western inference provider serving the same checkpoint — but it is a decision to make deliberately, not one to discover during an audit.

Do not rebuild on a promotion

The 50% discount ends on 9 September 2026. Build the business case on $0.15 and $0.50, treat the promotional period as a free evaluation window, and you will not be repricing a live system in the second week of September.

Watch the category, not the model

GLM-5.3-Flash is one of several open-weight releases landing in the same weeks, alongside Tencent’s Hy4 preview and Alibaba’s Qwen3.8-Flash. The pattern — large sparse mixture-of-experts, million-token context, permissive licence, aggressive price — is the story. Our AI models, tools and releases hub tracks these as they ship.

GLM-5.3-Flash: Frequently Asked Questions

Is GLM-5.3-Flash actually free?

It was free during the Ox Alpha preview week. It is not free now. The API lists at $0.15 per million input tokens and $0.50 per million output, currently halved under a promotion that ends on 9 September 2026. The weights themselves are free to download and use under an MIT licence.

Does GLM-5.3-Flash beat Claude Opus 4.8?

On some measures. It matches Opus 4.8’s Artificial Analysis Intelligence Index score of 57, beats it on DeepSWE v1.1 and AutomationBench, and trails it narrowly on Terminal-Bench 2.1 and on Humanity’s Last Exam with tools. “Comparable at a tenth of the cost” is a fairer summary than “better”.

Can I run GLM-5.3-Flash on my own machine?

Not on a workstation. The FP8 checkpoint is around 331 GB and needs roughly 386 GB of VRAM at tensor-parallel 4 on Hopper-generation hardware or newer. An eight-GPU H100 node is the practical entry point.

What does the Chinese chip claim really mean?

That the preview was served on roughly 100,000 domestically produced accelerators rather than Nvidia hardware. Z.ai has not named the vendor and says the chips have less compute and memory than Nvidia parts. It is evidence that export controls are not an absolute ceiling, not evidence of hardware parity.

Is the one-million-token context window usable?

It is servable, but the published evaluations were run at roughly 300,000 text tokens and 164,000 vision tokens. Treat anything beyond that as untested and validate it against your own workload before depending on it.

Which provider should I use for GLM-5.3-Flash?

It depends on whether you are optimising for latency or price, and the spread is wide — 47 to 273 tokens per second across providers, and $0.05 to $0.12 per million blended. Test two or three against your own traffic rather than taking the default route.

How does GLM-5.3-Flash compare with GLM-5.3?

Flash is the smaller, cheaper, faster-to-serve variant: 18 billion active parameters, about one-tenth the blended price, roughly three times less attention compute and a KV cache 4.44 times smaller. GLM-5.3 remains stronger on the hardest reasoning tasks.