GLM-5.3-Flash — chat with a 320B multimodal model, and a table it does not lead

Ten of the sixteen scored rows, and every one of them is printed by the same official table. On Terminal Bench 2.1 it scores 84.3 against GPT-5.6 Terra at 87.4, Gemini 3.7 Flash at 85.8 and Claude Opus 4.8 at 85.0. On DeepSWE v1.1 it scores 63.4 against 69.6 and 65.3. On NL2Repo it scores 56.3 against DeepSeek-V4-Vision-Exp at 57.7 and Claude Opus 4.8 at 69.7. On Agents’ Last Exam it scores 26.3 against 28.0, 27.3 and 27.0. On BabyVision, 53.4 against 70.9 and 61.6. On CharXiv Reasoning with tools, 89.4 against 89.9; on MVbench, 77.8 against 82.2; on MMVU, 80.5 against 82.3; on AutomationBench v1.0.6, 48.8 against 52.3; and on HLE with tools, 55.3 against 57.9. Counted across all sixteen rows the model finishes first six times, second five times, third three times and fourth twice, and it never finishes fifth. That distribution is why the comparison above is written against GLM-5.2 rather than as a claim about the whole table.

Ox-alpha is the alias under which the provider says it served this model before release. The launch post states that GLM-5.3-Flash was tested anonymously as ox-alpha on OpenCode and OpenRouter to gather user feedback, that it quickly became the most popular model of the week on those services, and that all of that traffic was served on Chinese AI chips. What the post does not give is any traffic figure, date range or comparison set, which is why this page reports the claim with its single source rather than as a measurement. The alias does not appear in the model code or in the API surface — the model code is glm-5.3-flash — and the practical use of knowing it is that third-party write-ups of this model often carry ox-alpha in the title, so articles under either name are discussing the same weights.

They are different models in the same line, released five weeks apart, and the official pages separate them on modalities first. GLM-5.3-Flash is described as the first native multimodal model in the GLM-5 series, accepting video, images, text and files, while the Chinese documentation page for GLM-5.3 states that it currently handles text modality only. Their documented limits are otherwise the same, at a 1M-token context window and a 128K maximum output, and both are quoted at a 128K output ceiling. On price they are not: the pricing page lists GLM-5.3-Flash at $0.15 per million input tokens and $0.50 per million output tokens, and GLM-5.3 at $1.4 and $4.4. Each is documented with its own architecture story, so this page reports the Flash model’s and names the flagship’s only where the two official pages overlap.

Images and PDFs, yes; video, no. The catalog entry behind the page publishes image among its input modalities, so JPEG, PNG and WebP files are accepted alongside PDFs, each capped at 24 MB. A PDF is converted to text before the model sees it, so a scan the extractor cannot read returns an error rather than an empty answer. Video is a different matter. Both the official page and the catalog entry behind this page list video among this model’s four input modalities, but the attachment adapter accepts only image types and PDF, so a video cannot be attached here in any form. That is a gap between what the model accepts and what this composer offers rather than a property of the model.

Because the ceiling is set by this workspace rather than by the model. The official pages document a 128K maximum output for the API, and the model page and the migration guide state the same figure. This site requests at most 4,096 output tokens per reply, and that number is a composer setting. It is worth being precise about where each number comes from, because the provider’s two documentation pages differ on a related value: the migration guide tells callers to set `max_tokens` appropriately against the 128K ceiling, while the model page’s own recommended-settings block lists sampling parameters and a reasoning effort instead. Neither of those is this site’s limit. Because the catalog entry behind this page publishes no context length either, the interface shows no context figure at all, so the official 1M-token window describes the API rather than this page. Long deliverables are best asked for in parts, carrying the context forward between them.

The official pricing page carries three model figures per million tokens in US dollars — $0.15 for input, $0.03 for cached input and $0.50 for output — with a cached-storage column reading “Limited-time Free” for which no end date is published. The dated offer lives on a separate official page and is about the subscription rather than the API: the GLM-5.3-Flash usage campaign in the GLM Coding Plan runs from 3 September 2026 to 7 October 2026, with the end date moved out from 20 September, during the daily window of 23:00 to 09:00 Singapore time. Inside that window, calls through ZCode or AutoClaw consume no quota, calls through other supported agents get double the plan’s standard quota, and the campaign applies to GLM-5.3-Flash only — selecting GLM-5.3 still consumes quota normally. One official statement is left out on purpose: the Chinese documentation page says this model is priced at one twentieth of GLM-5.3 during a limited-time discount, while the English pricing table carries only the one-tenth figures and the discount sentence has no date or matching row, so this page records the discrepancy instead of the lower number. This site charges points, with a catalog reference price of 7.8378 points per million input tokens and 27.027 per million output tokens.

Thinking cannot be disabled, and three reasoning levels are documented. The migration guide states that `thinking.type` accepts only `enabled`, that `reasoning_effort` chooses between low for lighter thinking, high for enhanced thinking and max for deep thinking, and that max is the default. The same guide treats this as a breaking change rather than a new option, and warns that applications currently sending `thinking.type: "disabled"` must switch that value before updating the model id or the request will fail. The model page repeats the constraint from the other direction, describing the capability as supporting multiple thinking modes while stating that thinking cannot be disabled. Sampling parameters are documented alongside it: temperature defaults to 1 and top_p to 0.95, and the guide advises adjusting only one of the two. None of this is controllable in this chat box — the catalog entry marks reasoning as required and publishes no level list, so the composer offers neither a level selector nor an off switch.

The launch post gives three changes and attaches a number to each. The architecture is a hybrid of linear attention, which captures local dependencies through state modelling, and sparse attention, which retrieves relevant global context through a lightweight indexer; the provider describes it as the first open-source frontier model to combine the two. IndexPool compresses four indexer key vectors into one by weighted pooling, which the post introduces specifically to reduce the indexer’s latency and memory overhead at a 1M-token context, and the models the provider reports against are GLM-5.3, DeepSeek-V4-Flash and Kimi-K3 with attention compute and KV cache normalised per head per layer so that differently sized models can be compared: 3.0× less attention compute and a 4.4× smaller KV cache than GLM-5.3, and a KV cache still slightly larger than the two smaller rivals. Manifold-Constrained Hyper-Connections are credited with better scaling efficiency, and the post attributes part of the gain to a 30T-token multimodal pre-training corpus. Serving is the third part: the model runs on a large cluster of Chinese accelerators through an Encode–Prefill–Decode disaggregated architecture, with the provider reporting a 3× end-to-end improvement over its initial baseline on the same hardware.