GLM-5.3 Flash Review — Zhipu's First Multimodal GLM at Flash Prices
On August 26, 2026, Zhipu AI (Z.ai) released GLM-5.3 Flash, and it is a different kind of GLM drop from the ones that came before. Every GLM-5 release to date — GLM-5.3 two weeks earlier, GLM-5.2, GLM-5.1 — was a text-and-code model. Flash is the first natively multimodal model in the line, taking image and video input, and it ships at roughly a tenth of the full GLM-5.3 price under the same unrestricted MIT license. This review covers what GLM-5.3 Flash is, which of its numbers are measured versus estimated, and where a cheap multimodal open model actually fits.
TL;DR verdict
| GLM-5.3 Flash | |
|---|---|
| Type | Multimodal general LLM (flash tier) |
| Architecture | Sparse MoE, ~18B active / ~320B total |
| Modalities | Text, image, video input |
| Context window | ~1M tokens · up to 131,072 output |
| License | MIT (unrestricted open weights) |
| Pricing | $0.15 in / $0.50 out per 1M · launch promo $0.075/$0.25 |
| Headline number | Artificial Analysis Intelligence Index 57 · DeepSWE 63.4 |
| Availability | Hugging Face weights, Z.ai API, OpenRouter |
| Best for | High-volume multimodal inference, cheap self-hosted vision workloads |
| Caveat | Published an index score and a coding number, not the full standard suite |
If you skip the rest: GLM-5.3 Flash brings native image and video input to the GLM-5 line for the first time, at a price that undercuts almost every other multimodal model on the board. It reports an Artificial Analysis Intelligence Index of 57 — well above the median for models in its price class — and a DeepSWE score of 63.4, a large jump over GLM-5.2’s 46.2, all under a permissive MIT license. The honest asterisk is the same one that followed GLM-5.2: Zhipu led with an index number and a coding figure, not the full standard suite, so its cross-vendor placements are still interpolation.
What it is
GLM-5.3 Flash is a sparse Mixture-of-Experts model with roughly 320 billion total parameters and ~18 billion active per token. That low active-parameter count is the whole point of a “flash” tier: it keeps inference cheap and fast enough to justify the pricing, the same lever the Gemini 3.7 Flash and DeepSeek V4 Flash releases pulled in their own tiers. It is positioned below the full GLM-5.3, which runs on the larger 743B base, rather than as its replacement.
The headline architectural change is native multimodality. GLM-5.3 Flash accepts image and video input alongside text — the first GLM-5 model to do so — which pulls the line into the same multimodal bracket the closed flash tiers have occupied for a while. The context window is about 1 million tokens with up to 131,072 tokens of output, enough to hold a long document set or a running agentic session. And the license is, again, the part that matters most for a lot of teams: unrestricted MIT, weights downloadable from Hugging Face, so you can self-host, fine-tune, and run it air-gapped.
What the launch numbers say
Z.ai led the launch with two figures rather than a full benchmark card:
| Metric | GLM-5.3 Flash | Reference | What it measures |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 57 | median for the class ≈ 29 | Composite of independent evaluations |
| DeepSWE | 63.4 | GLM-5.2: 46.2 | Long-horizon software engineering |
The Artificial Analysis Intelligence Index of 57 is an independent composite, which makes it a more useful single number than a proprietary in-house score — it places Flash well above the median for models in its price band. The DeepSWE 63.4 is vendor-reported, but the anchor is credible: GLM-5.2 scored 46.2 on the same benchmark, so a jump to 63.4 reflects the coding-focused post-training Zhipu has been pushing across the 5.2-to-5.3 generation rather than an implausible leap.
What is not in the launch material: GPQA Diamond, SWE-bench Pro, HLE — the standard columns our leaderboard tracks. That is why, in our models leaderboard, GLM-5.3 Flash’s benchmark cells are marked as conservative estimates, anchored to the full GLM-5.3 and to GLM-5.2 and discounted for the smaller Flash active-parameter budget. Treat those cells as directional placements until independent runs land, exactly as we did for GLM-5.2 at launch.
What it costs
On Zhipu’s first-party API, the launch list price is:
- Input: ~$0.15 per 1M tokens
- Output: ~$0.50 per 1M tokens
- Cached input: ~$0.03 per 1M tokens
A 50% launch promotion runs input at $0.075 and output at $0.25 per 1M tokens through September 9, 2026. Either way, this is roughly a tenth of the full GLM-5.3 price ($1.40/$4.40) and undercuts most multimodal flash tiers on the board — the closest closed comparison, a cheap multimodal flash model, still tends to list several times higher on output. And because the weights are MIT-licensed, the metered API is a floor, not a ceiling on savings: run your own GPUs and the per-token cost disappears entirely. The calculator on the leaderboard lets you plug in your own token mix against the closed flash tiers.
How it compares
GLM-5.3 Flash’s real competition is the cheap multimodal tier, not the absolute frontier. Against Gemini 3.7 Flash, Google’s flash multimodal model, GLM-5.3 Flash trades away polish and ecosystem for open weights and a lower floor on cost — and unlike Gemini it can be self-hosted. Against DeepSeek V4 Flash, the two occupy the same “frontier-adjacent open weights, priced to move” niche, but Flash’s native video input is a capability DeepSeek’s flash tier does not match. And against the full GLM-5.3 it is the cheaper, lighter, multimodal sibling: you give up raw reasoning headroom in exchange for vision, video, and a roughly 10× cost cut.
Against closed leaders like Claude Opus 5 or the GPT-5.6 tiers, GLM-5.3 Flash is not trying to compete on the hardest reasoning — it is trying to make multimodal inference cheap enough to run at volume. That is the same pattern the whole open tier has followed for months: close the gap where it is cheapest to close (throughput, price, now multimodality) while the closed leaders keep the reasoning crown.
Who should care
- Teams running high-volume multimodal pipelines: This is the headline use case. Image and video input, a 1M context window, MIT weights, and a price low enough to run at scale. Pull it from Hugging Face and A/B it against your current vision model.
- Cost-sensitive agentic workloads: If you are paying closed-model flash rates on a high-volume job, GLM-5.3 Flash’s roughly 10× cost advantage over the full GLM-5.3 plus the self-host option is worth a pilot — see multi-agent pipelines for where a cheap, capable model slots into a larger workflow.
- Anyone benchmarking open vs closed multimodal: Flash is a useful data point because it reports an independent index score, not just proprietary numbers. Still, confirm its standard-suite placements with third-party runs before ranking it in production planning.
- Teams that need the hardest reasoning: Look higher up the board. Flash is a value multimodal model; for frontier reasoning a full-size model is the better fit — the LLM Benchmark Comparison 2026 covers how to weigh that trade.
FAQ
What is GLM-5.3 Flash? Zhipu AI’s flash-tier, natively multimodal model, released August 26, 2026. A sparse MoE (~18B active, ~320B total) with image and video input, a context window of about 1M tokens, up to 131,072 tokens of output, and an unrestricted MIT open-weights license.
How much does it cost? About $0.15 per 1M input tokens and $0.50 per 1M output on the Z.ai API, with a launch promo of $0.075/$0.25 through September 9, 2026. MIT weights mean self-hosting is also an option.
Is it multimodal? Yes — it is the first GLM-5 model to take image and video input. GLM-5.3, GLM-5.2, and GLM-5.1 were text and code only.
What benchmarks did Zhipu publish? An Artificial Analysis Intelligence Index of 57 and a DeepSWE score of 63.4 (up from GLM-5.2’s 46.2). It did not publish the full standard suite (GPQA, SWE-bench Pro, HLE), so those remain estimates anchored to GLM-5.3 and GLM-5.2.
Can I self-host it? Yes — the weights are on Hugging Face under an unrestricted MIT license, so you can run, fine-tune, and deploy it on your own infrastructure.
Continue reading
- AI Models Leaderboard — GLM-5.3 Flash versus 60+ models on benchmarks, pricing, and context window, with a cost calculator.
- GLM-5.2 Review — the open-weight coding flagship this Flash tier descends from, and how Zhipu’s benchmark transparency has evolved.
- Gemini 3.7 Flash Review — the closed multimodal flash tier GLM-5.3 Flash is priced against.
- DeepSeek V4 Flash 0731 Review — another open flash-tier release in the same value bracket.
- LLM Benchmark Comparison 2026 — how to read index scores and self-reported numbers without getting fooled.
- All Reviews — index of every head-to-head review on the site.