Quick answer: GLM 5.3 Flash is Z.ai's efficiency-tier model in the GLM-5 family, released August 26, 2026. It is a 320B-parameter Mixture-of-Experts model with 18B active parameters, the first natively multimodal model in the series (text + image + video in, text out), with a 1,048,576-token context window, open weights under the MIT license, and a list API price of $0.15 per 1M input / $0.50 per 1M output tokens. It is also the model that spent its first week on OpenRouter anonymously as the stealth model "Ox Alpha."
If you tried to research this model in the past week, you probably hit the same mess everyone else did. Half the pages you found were about GLM 5.3 (a different, more expensive model), half were about a mystery endpoint called Ox Alpha, and the pricing numbers disagreed with each other by a factor of two. None of that is your fault — the model shipped under a codename, then got revealed, and then launched with a temporary discount that made every price quote instantly ambiguous.
This article is built from the primary sources only: the Hugging Face model card, the zai-org/GLM-5 repository, the OpenRouter model page and its public endpoints API (queried August 27, 2026), and Artificial Analysis's independent measurements. Where a number comes from a community report rather than a vendor, it is labelled as such. We have not run our own benchmark suite against GLM 5.3 Flash, and we say so rather than dressing up vendor charts as testing.
What This Article Solves
The pain point is name collision plus price ambiguity. "GLM 5.3" and "GLM 5.3 Flash" are two different models with roughly a 10x price gap between them, and the Flash launch price you see quoted almost everywhere is a temporary 50% discount, not the list rate. On top of that, the model was previewed anonymously, so a large share of the community knowledge about it is filed under a name that no longer exists.
By the end of this page you will know exactly what GLM 5.3 Flash is architecturally, what it actually costs at list versus at discount, where its benchmark numbers come from and which ones are independent, and — the part most explainers skip — the specific workloads where a 18B-active model is the wrong pick.
GLM 5.3 Flash at a Glance
| Field | Value |
|---|---|
| Full name | GLM-5.3-Flash |
| Vendor | Z.ai (Zhipu AI) |
| Released | August 26, 2026 |
| Previously known as | Ox Alpha (stealth/ox-alpha, preview from August 20, 2026) |
| Total parameters | 320B |
| Active parameters | 18B per token |
| Architecture | Mixture-of-Experts, 45 layers, 8 of 288 experts routed per token |
| Attention | Hybrid — KDA linear-attention layers + NoPE sparse MLA layers |
| Context window | 1,048,576 tokens (1M) |
| Max output | 131,072 tokens |
| Modality | text + image + video → text |
| Weight precision | Native FP8 (a BF16 release also exists) |
| Training corpus | 30T-token multimodal pre-training corpus |
| Weights license | MIT |
| List API price | $0.15 / 1M input, $0.03 / 1M cached input, $0.50 / 1M output |
The Architecture, in Plain Terms
The number that matters most is not 320B — it is 18B. GLM 5.3 Flash is a Mixture-of-Experts model, so a given token only activates 8 of the 288 experts. You pay storage and memory for 320B parameters, but you pay compute closer to an 18B dense model. That single fact explains the model's entire personality: frontier-adjacent quality at a price point that looks like a mistake.
Three design choices are worth understanding because they show up directly in how the model behaves:
1. Hybrid sparse + linear attention. This is the first time the GLM series has shipped a hybrid attention stack: KDA linear-attention layers handle local dependencies, while NoPE sparse MLA layers pull in the globally relevant context. Z.ai reports this delivers roughly 3x less attention compute and a 4.4x smaller KV cache than GLM-5.3. In practice, that smaller KV cache is why a million-token context is affordable here rather than theoretical — long-context pricing is usually strangled by KV memory, not by the weights.
2. IndexPool. To keep the million-token window usable rather than just advertised, the model compresses indexer key vectors through weighted pooling, trading a little precision for meaningfully lower latency and memory at long context.
3. mHC (Manifold-Constrained Hyper-Connections). Z.ai credits this with letting them halve both active parameters and layer count relative to GLM-4.5 at a comparable total parameter count. It is the scaling-efficiency trick that makes the 320B/18B ratio work.
You do not need to internalise any of this to use the model. You do need it to explain to a colleague why a "320B model" is being served at $0.15 per million input tokens without anyone losing money.
Natively Multimodal — and Why "Natively" Matters
GLM 5.3 Flash is the first natively multimodal model in the GLM-5 family. It accepts text, images, and video as input and returns text. "Natively" means the visual pathway is part of the pre-training corpus (30T tokens, multimodal) rather than a vision encoder bolted onto a text model afterwards.
The practical consequence: tasks that mix code and visual context — reading a screenshot of a broken UI and patching the CSS, tracing a bug from a screen recording, extracting a schema from a diagram — are in the model's core competence rather than at the edge of a retrofit. Z.ai's own OfficeQA Pro number (62.4, which it reports as ahead of Claude Opus 4.8) is a document-and-visual reasoning benchmark, and it is the one place where the Flash tier claims an outright lead over a flagship rather than a near-miss.
If you want to check that claim against your own screenshots rather than a chart, you can open a GLM 5.3 Flash session on glm5.app and paste one in — a five-minute test tells you more about fit than any benchmark table.
What It Costs (Read the Footnote)
The list price from Z.ai is $0.15 per 1M input tokens, $0.03 per 1M cached input, and $0.50 per 1M output tokens.
The number you will see quoted almost everywhere right now — $0.075 input / $0.25 output — is that list price with a launch 50% discount applied, and it is currently what Z.AI's own OpenRouter endpoint, Novita, and GMICloud charge. Other providers on the same model (DeepInfra, Baseten, Parasail, Io Net, Cloudflare) are already serving at the undiscounted $0.15 / $0.50.
Two rules follow, and they are the difference between a budget that holds and one that doubles overnight:
- Model your forecast at $0.15 / $0.50. The discount is promotional. If your unit economics only work at the discounted rate, you do not have unit economics.
- Name the provider when you quote a price. "GLM 5.3 Flash costs $0.075" is only true on three of the ten endpoints serving it.
The full ten-provider spread is broken down in our GLM 5.3 Flash pricing guide. Independent tracker Artificial Analysis lists a blended price of $0.10 per 1M tokens using a 7:2:1 cache-hit / input / output mix — a useful single number for comparison shopping, but only if your traffic actually looks like that ratio.
Benchmarks: What Is Vendor-Reported vs. Independent
Z.ai's own reported figures, with the comparisons it chose:
| Benchmark | GLM 5.3 Flash | Reference |
|---|---|---|
| Terminal-Bench 2.1 | 84.3 | Claude Opus 4.8: 85.0 · GPT-5.6 Terra: 87.4 |
| DeepSWE v1.1 | 63.4 | GLM-5.2: 46.2 |
| AutomationBench | 48.8 | GLM-5.2: 26.2 |
| Z.ai Code Bench v1.0 (max effort) | 29.0 | Claude Opus 4.8: 29.5 |
| OfficeQA Pro | 62.4 | reported ahead of Opus 4.8 |
| HLE (with tools) | 55.3 | — |
Treat the first-party table with the usual caution — "Z.ai Code Bench" is Z.ai's own benchmark, and a vendor picking the comparison set is a vendor picking the comparison set. The genuinely useful signal is the GLM-5.2 delta: DeepSWE v1.1 goes from 46.2 to 63.4 and AutomationBench from 26.2 to 48.8, against a model that costs roughly ten times more. That is a within-family, same-methodology comparison, which is the kind that survives scrutiny.
Independently, Artificial Analysis scores GLM 5.3 Flash at 57 on its Intelligence Index, well above the ~27 median for open-weight models of similar size. The same source measures output at 50.2 tokens/second (below the 67 t/s median for its class — this model is not fast in the throughput sense) with a 1.47s time to first token (better than the 2.14s median).
That combination is worth stating plainly, because the name misleads: "Flash" here means cheap, not fast. It is a low-cost tier, not a low-latency tier. If your product needs tokens on screen quickly, benchmark it before you assume.
Open Weights: MIT, and What That Buys You
The weights are on Hugging Face at zai-org/GLM-5.3-Flash (with a -BF16 variant) and on ModelScope, released under the MIT license — about as permissive as open weights get. Commercial use, modification, redistribution, and private forks are all allowed with no revenue threshold and no acceptable-use appendix. Note one wrinkle worth getting right in a compliance review: the weights are MIT, while the zai-org/GLM-5 code repository carries Apache-2.0.
There is no standalone GLM-5.3-Flash GitHub repository. The inference recipes and architecture notes live in zai-org/GLM-5, which is the repo to point your team at. Officially supported serving stacks are SGLang, vLLM, TokenSpeed, and KTransformers, and the community has already produced GGUF quantizations for llama.cpp, Ollama, LM Studio, and Jan.
Full architecture and download details are in our GLM 5.3 Flash parameters and size breakdown. Self-hosting is real but not casual: the FP8 weights total roughly 328 GB, per file-size readings reported on NVIDIA's developer forums. That is a multi-GPU-node conversation, not a workstation one.
The Ox Alpha Connection
For six days, GLM 5.3 Flash existed on OpenRouter as stealth/ox-alpha — an anonymous, free-to-use reasoning model with no disclosed provider. The full before/after — model ID, pricing, data policy — is covered in Ox Alpha vs GLM 5.3 Flash. Z.ai's launch announcement confirmed the link directly, noting the model had previously been previewed as Ox Alpha and, notably, that it had been running entirely on Chinese AI chips during that preview. Reporting around the launch describes a custom SGLang-derived serving engine delivering a roughly 3x end-to-end serving improvement; the specific chip vendor has not been disclosed.
This matters for two reasons beyond trivia. First, a large body of practical community experience — throughput observations, agentic-coding results, quirks — is filed under "Ox Alpha," so that is the search term to use when you want field reports rather than launch-day coverage. Second, the preview period generated real production traffic (OpenRouter's own activity panel showed the top five consumers were all agentic coding tools), which means the launch numbers rest on a week of genuine load rather than a cold start.
When GLM 5.3 Flash Is the Wrong Choice
Most explainers stop at the spec sheet. Here is the part that saves you a wasted sprint — three situations where this model is a bad fit:
You need low latency, not low cost. At ~50 tokens/second measured output, GLM 5.3 Flash is slower than most of its price peers. For a streaming chat UI where perceived speed is the product, a faster model at a higher price is often the cheaper decision.
Your task lives at the frontier's edge. The gap to Claude Opus 4.8 is small on Z.ai's chosen benchmarks (84.3 vs 85.0 on Terminal-Bench 2.1) but small gaps on benchmarks become large gaps on the hardest 5% of real tasks. For the specific jobs where a wrong answer is expensive to detect, the flagship premium is usually justified.
You want to self-host on one box. 328 GB of FP8 weights does not fit a single 128 GB machine. Community reports on NVIDIA's forums put the practical floor at multiple clustered units with 4-bit or hybrid quantization — and quantizing away a third of the precision changes the quality picture the benchmarks describe.
Where it is the right choice is unambiguous: high-volume agentic coding, long-context document and repository work, and any pipeline where you were previously paying GLM 5.2 or a flagship rate for tasks that did not need flagship depth. The honest way to find out is to run your own hardest real task through it — try GLM 5.3 Flash on glm5.app with no setup, or wire the glm-5.3-flash model ID into your existing pipeline and diff the outputs against whatever you run today.
Frequently Asked Questions
Is GLM 5.3 Flash the same as GLM 5.3? No. They are different models at different price tiers. GLM 5.3 Flash is the 320B-A18B multimodal efficiency model released August 26, 2026; GLM 5.3 is the more expensive, text-focused release from earlier in August. Z.ai reports Flash beats GLM-5.2 across benchmarks at roughly one-tenth the price, and Coding Plan users get about 3x the usable quota on Flash compared to GLM-5.3.
Is GLM 5.3 Flash free? Not permanently. It was free during its Ox Alpha preview week, and the current $0.075 / $0.25 rate is a temporary 50% launch discount off the $0.15 / $0.50 list price. The weights are free in the open-source sense — MIT-licensed on Hugging Face — but running them yourself means supplying ~328 GB of GPU memory.
Can it really handle a million tokens? The context window is 1,048,576 tokens on Z.ai's own endpoint and most OpenRouter providers. Two caveats: Io Net serves a 262,144-token variant and Cloudflare lists 1,310,720, so the ceiling depends on which endpoint you hit; and max output is capped separately at 131,072 tokens on most endpoints.
What hardware do I need to run it locally? More than one box. The FP8 release is roughly 328 GB. Community discussion on NVIDIA's DGX Spark forums centres on 2x units with 4-bit or hybrid quantization as the realistic entry point, with reported peak throughput around 43 tok/s on that setup — community-measured, not independently verified.
Where do I get the weights?
zai-org/GLM-5.3-Flash on Hugging Face (or zai-org/GLM-5.3-Flash-BF16 for BF16), and ZhipuAI/GLM-5.3-Flash on ModelScope. There is no dedicated GitHub repo — inference recipes live in zai-org/GLM-5.
Sources
- zai-org/GLM-5.3-Flash — Hugging Face model card — parameters, architecture, MIT license, benchmark table, quantization availability.
- zai-org/GLM-5 — GitHub — official hybrid-attention description, FP8/BF16 variants, supported inference frameworks.
- GLM 5.3 Flash on OpenRouter — release date, context window, per-provider pricing, Ox Alpha lineage.
- OpenRouter endpoints API for z-ai/glm-5.3-flash — the per-provider price and context table, queried August 27, 2026.
- Artificial Analysis — GLM-5.3-Flash — independent Intelligence Index, output speed, TTFT, and blended price.
Prices and endpoint availability change frequently. All figures were verified on August 27, 2026; confirm against the OpenRouter model page or Z.ai's pricing documentation before committing a production budget.
