GLM-5.3-Flash is the August 26, 2026 release that puts Z.ai’s cost argument in one number: $0.15 per million input tokens. It is the first natively multimodal model in the GLM-5 series, trained from a new base — 320B total parameters, 18B active — rather than post-trained from GLM-5.2’s.
The architecture is the interesting part. It combines linear attention (local dependencies via state) with sparse attention (global retrieval through a lightweight indexer, with a new “IndexPool” that compresses four indexer keys into one), plus Manifold-Constrained Hyper-Connections. Against GLM-5.3 that means 3.0x less attention compute and a 4.4x smaller KV cache — which is how a 1M-token context becomes affordable. Weights are on Hugging Face under the MIT license.
Benchmarks
Vendor-reported by Z.ai at launch (August 26, 2026). Comparison columns are as published.
| Benchmark | GLM-5.3-Flash | GLM-5.2 | DeepSeek V4-Vision-Exp | Opus 4.8 | GPT-5.6 Terra | Gemini 3.7 Flash |
|---|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 84.3 | 81.0 | 83.9 | 85.0 | 87.4 | 85.8 |
| DeepSWE v1.1 | 63.4 | 46.2 | 59.3 | 58.0 | 69.6 | 65.3 |
| NL2Repo | 56.3 | 48.9 | 57.7 | 69.7 | — | — |
| Toolathlon Verified | 78.4 | 59.9 | 75.9 | 76.2 | 74.9 | — |
| AutomationBench | 48.8 | 26.2 | 38.8 | 41.0 | 37.2 | 52.3 |
| Agents’ Last Exam | 26.3 | 20.4 | 27.3 | 27.0 | 28.0 | — |
| HLE with tools | 55.3 | 54.7 | 55.1 | 57.9 | — | — |
| GDPval-AA v2 (Elo) | 1773 | 1504 | 1675 | 1582 | 1571 | 1527 |
| OSWorld 2.0 | 59.1 | — | 54.9 | 54.8 | 50.2 | 47.9 |
| Vision2Web | 77.8 | — | 67.6 | 76.1 | 67.5 | 76.5 |
| CharXiv Reasoning (tools) | 89.4 | — | 80.4 | 89.9 | 88.0 | 88.7 |
| MMVU | 80.5 | — | 72.7 | 67.4 | 75.8 | 82.3 |
Opus 4.8 and GPT-5.6 Terra are older-generation comparison points; Z.ai’s table does not include the newer Claude 5.5 or GPT-6 models. Gemini 3.7 Flash beats GLM-5.3-Flash on AutomationBench, MMVU and BabyVision (70.9 vs 53.4).
On Z.ai’s private Code Bench v1.0, it nearly matches Opus 4.8 at max effort (29.0 vs 29.5). Z.ai also cites Artificial Analysis’ Intelligence Index v4.1.1, where it scores 57 at about $0.045 per task at discounted prices.
Pricing
| Per 1M tokens | GLM-5.3-Flash | GLM-5.3-FlashX |
|---|---|---|
| Input | $0.15 | $0.37 |
| Cached input | $0.03 | $0.075 |
| Output | $0.50 | $1.25 |
FlashX is the faster tier (about 200 tokens per second); both support the 1M-token context.
Specs
| Spec | Value |
|---|---|
| API models | glm-5.3-flash, glm-5.3-flashx |
| Architecture | 320B MoE, 18B active, 45 layers; hybrid sparse and linear attention, IndexPool, mHC |
| Pre-training | New base, 30T multimodal tokens |
| Context window | 1M tokens; max output 128K |
| Modalities | Text, image, video and file in; text out |
| Reasoning | Always on; recommended reasoning_effort: max, temperature 1, top_p 0.95 |
| License | MIT (weights on Hugging Face) |
| Serving | SGLang, vLLM, TokenSpeed; served at scale on Chinese AI chips at launch |
How it compares
- vs GLM-5.3: the flagship is bigger, text-only and roughly 9x more expensive; Flash wins on price, vision and long-context cost.
- vs GLM-5.2: better on every published row, at roughly one-tenth the price.
- vs DeepSeek V4.1-Flash: both are efficient open models; DeepSeek’s is larger (552B) and scores higher on DeepSWE v1.1 (74.2 vs 63.4), while GLM-5.3-Flash has native video input.
- vs Qwen3.8-Flash-Next: a smaller open model (6B active) at a similar API price; Alibaba reports 58.7 on DeepSWE v1.1 against GLM-5.3-Flash’s 63.4 (different harnesses), and ships it under a community license rather than MIT.
Limitations
- Vendor-reported, with older comparison models.
- Always-on thinking — no non-reasoning mode.
- Text output only.
For the wider picture, see the 2026 LLM guide.
Best for
- Low-cost coding agents with visual feedback loops (frontend, games, GUIs)
- Computer-use and browser-use agents
- Office and document workflows — PPTX, DOCX, XLSX and PDF deliverables
- Self-hosted multimodal agent stacks under an MIT license
Pros & cons
Strengths
- Priced at $0.15 / $0.50 — roughly one-tenth of GLM-5.2 — yet beats it on every coding and agentic row Z.ai published
- First natively multimodal GLM-5 model, taking images, video and files, with strong visual-coding and computer-use scores (OSWorld 2.0 59.1, Vision2Web 77.8)
- A new hybrid sparse-plus-linear attention design cuts attention compute 3.0x and KV cache 4.4x versus GLM-5.3, making 1M context cheap to serve
- MIT-licensed weights on Hugging Face; Z.ai cites a 57 on Artificial Analysis' Intelligence Index v4.1.1 at about $0.045 per task
Limitations
- Still behind the biggest models on hard long-horizon tasks, and Z.ai itself says it only "approaches" Claude Opus 4.8
- A 320B-parameter model is a multi-GPU deployment, though far lighter than GLM-5.3
- Thinking cannot be disabled, as with GLM-5.3, so latency-sensitive uses must pick low effort
- Output is text only — no image or audio generation
Frequently asked questions
When was GLM-5.3-Flash released?
Z.ai released it on August 26, 2026, with weights on Hugging Face and API access as `glm-5.3-flash` and `glm-5.3-flashx`. Before launch it was tested anonymously as "ox-alpha" on OpenCode and OpenRouter.
How much does GLM-5.3-Flash cost?
GLM-5.3-Flash is $0.15 per 1M input tokens, $0.03 cached and $0.50 output. GLM-5.3-FlashX is $0.37 input, $0.075 cached and $1.25 output, with faster inference of about 200 tokens per second.
Is it a distillation of GLM-5.3?
No. Z.ai says it starts from a newly trained base model with a redesigned architecture and a 30T-token multimodal pre-training corpus, not from the GLM-5.2 or GLM-5.3 base.
How big is it?
320B total parameters with 18B active across 45 layers, against 355B / 32B / 92 layers for GLM-4.5. Z.ai calls it the first open frontier-class model to combine sparse and linear attention.
Can GLM-5.3-Flash see images and video?
Yes. It accepts text, images, video and files and outputs text, with a 1M-token context window and 128K maximum output.