GLM-5.3-Flash (and GLM-5.3-FlashX)

2026-08-26 · Publisher : Z.ai · Domain : Artificial Intelligence

Z.ai released GLM-5.3-Flash, a natively multimodal MoE with 320B parameters (18B active), hybrid attention and 1M context, under the MIT license, priced at USD 0.15 input and USD 0.50 output per million tokens.

### Context GLM-5.2 and 5.3 have a 744B base and pricing of USD 1.4/4.4. GLM-5.3-Flash starts from a newly trained base, with the goal of a very low inference cost. According to the post, the model was tested anonymously before its release on OpenCode and OpenRouter, where it reportedly became the most popular model of the week, on Chinese chips. ### What the model brings - A hybrid architecture that combines linear attention (local dependencies) and sparse attention with a lightweight indexer; IndexPool compresses four indexer key vectors into one. - Manifold-Constrained Hyper-Connections (mHC) for scaling efficiency. - Compared with GLM-5.3, 3.0 times less attention compute and a 4.4 times smaller KV cache; 45 layers vs 92 for GLM-4.5, and 18B active vs 32B. - The first natively multimodal GLM (image, video, text, files), with visual use in the coding loop (self-checking the rendering) and interface interaction (OSWorld 2.0: 59.1). - A score of 57 on the Artificial Analysis Intelligence Index v4.1.1 for USD 0.045 per task (discounted pricing), according to the vendor. ### Architecture and specifications - 320B in total, 18B active (about 321.3 billion by the HF count). Multimodal pretraining corpus of 30T tokens. - Context: 1M tokens; maximum output: 128K; inputs: image, video, text, file; output: text (API documentation). Knowledge cutoff: not disclosed. - License: MIT (LICENSE file, "Copyright (c) 2026 Z.AI Co., Ltd"). Weights: zai-org/GLM-5.3-Flash and BF16 version. - reasoning_effort parameter: low, high, max (max by default). ### Cost and access | Model (Z.ai API, current pricing) | Input /M | Cached input /M | Output /M | |---|---|---|---| | GLM-5.3-Flash | USD 0.15 | USD 0.03 | USD 0.50 | | GLM-5.3-FlashX (about 200 tokens/s) | USD 0.37 | USD 0.075 | USD 1.25 | Self-hosting: SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth; required hardware not quantified. ### Benchmarks | Benchmark | GLM-5.3-Flash | GLM-5.2 | Claude Opus 4.8 | GPT-5.6 Terra | Gemini 3.7 Flash | |---|---|---|---|---|---| | Terminal Bench 2.1 | 84.3 | 81.0 | 85.0 | 87.4 | 85.8 | | DeepSWE (v1.1) | 63.4 | 46.2 | 58 | 69.6 | 65.3 | | Toolathlon Verified | 78.4 | 59.9 | 76.2 | 74.9 | - | | AutomationBench (v1.0.6) | 48.8 | 26.2 | 41.0 | 37.2 | 52.3 | | HLE with tools | 55.3 | 54.7 | 57.9 | - | - | | GDPval-AA v2 | 1773 | 1504 | 1582 | 1571 | 1527 | | OSWorld 2.0 | 59.1 | - | 54.8 | 50.2 | 47.9 | | BabyVision | 53.4 | - | 46.8 | 61.6 | 70.9 | Source: z.ai post for GLM-5.3-Flash, scores published by the vendor; GDPval-AA v2 measured by Artificial Analysis. ### Versus closed models In the vendor's table, GLM-5.3-Flash trails Opus 4.8 on Terminal Bench 2.1 (84.3 vs 85.0) but is ahead on DeepSWE (63.4 vs 58), Toolathlon Verified (78.4 vs 76.2), AutomationBench (48.8 vs 41.0) and GDPval-AA v2 (1773 vs 1582). It is behind GPT-5.6 Terra on DeepSWE (69.6) and on Terminal Bench 2.1 (87.4), and behind Gemini 3.7 Flash on BabyVision (70.9 vs 53.4), so pure vision is not its strong point. At USD 0.15 per million input tokens, the price is about nine times lower than that of GLM-5.3 (USD 1.4), a gap consistent with the post's mention of "one tenth of the price"; no closed pricing has been verified. The model is under MIT, unlike GLM-5.3. ### Key takeaways GLM-5.3-Flash brings a new, smaller, multimodal base with a 1M context, with agent scores close to a frontier closed model for a very low inference cost. It remains under the MIT license while GLM-5.3 changed license.

Tags : LLM, Open Source, Z.ai, Zhipu AI, GLM, MIT, MoE, Multimodal