Today, Zhi Spectrum (02513) launched and open-sourced GLM-5.3-Flash (320B-A18B), the first native multi-modal model in the GLM-5 series.
People seem to have become accustomed to binding cutting-edge intelligence to cutting-edge prices, and think this is the cost that technological progress must pay. Today, GLM-5.3-Flash is here: 320B's overall parameters, surpassing GLM-5.2, and scored 57 points in the world's authoritative Artificial Analysis Intelligence Index (AA Comprehensive Intelligence Index), entering the global leading-edge model capability range, on par with Anthropic's most popular model Claude Opus 4.8. In the self-developed Z.ai Code Bench somatosensory evaluation, its programming performance was comparable to Claude Opus 4.8.
Meanwhile, GLM-5.3-Flash is priced at 1/10 of GLM-5.3, 1/20 of GLM-5.3 and 1/40 of Opus 4.8 during the limited-time discount. Same intelligence, 1/40 price, cutting-edge intelligence, no need to save money for the first time.
In order to gather extensive and professional feedback from a wide range of users, we conducted large-scale tests on OpenCode and OpenRouter using the anonymous model Ox-Alpha (known in the Chinese community as Niulai) before the official release. The Ox-Alpha quickly became the most popular model of the week, setting a new high in dual-platform calls. However, all of these requested traffic is supported by domestic chips with computing power.
In various benchmarks and actual use, GLM-5.3-Flash completely surpassed GLM-5.2, which had twice the number of parameters, but the price was only 1/10 of the latter. This is due to its new model architecture. The GLM-5.3-Flash architecture is designed for extremely low cost. The total number of parameters is comparable to GLM-4.5 (355B vs. 320B), but the amount of activation parameters (32B → 18B) and the number of layers (92 → 45) are almost halved. Combined with our latest 30T Token multi-modal pre-training corpus, GLM-5.3-Flash can achieve greater performance with fewer computing resources.
GLM-5.3-Flash has undergone multiple architecture upgrades at the same time. It is the first open source cutting-edge model that uses a hybrid architecture of sparse attention and linear attention. While maintaining accurate long-term context capabilities, it greatly reduces the cost of long context services; and uses Manifold-Constrained Hyper-Connections (MHC) to improve model scaling capabilities. See blog: https://z.ai/blog/glm-5.3-flash for technical details.
Extremely low cost architecture
To reduce the cost of attention in the long context scenarios of the model, we used a hybrid architecture combining linear attention and sparse attention. Among them, linear attention captures local dependencies through a recursive mechanism, while sparse attention uses a lightweight indexer to recall global context. We introduced IndexPool to further reduce the indexer's latency and memory overhead in the 1M context, and compressed the indexer's 4 cache vectors into 1 through weighted pooling.
To show the efficiency of the architecture, we compared GLM-5.3-Flash with GLM-5.3, Deepseek-v4-Flash, and Kimi-K3, focusing on the computational volume of a single token and the size of the KV cache. To fairly compare models of different sizes, we separately show the amount of attention computation for each head, each layer, and the average KV cache size (BF16) per layer. Compared with GLM-5.3, GLM-5.3-Flash's attention calculation volume and KV cache size were reduced by 3.01 times and 4.44 times, respectively.
Of all the baseline models, GLM-5.3-Flash has the least amount of attention calculations. However, its KV cache size is still slightly higher than Kimi-K3 and Deepseek-v4-Flash, which is our direction for further optimization in subsequent work.
Visual feedback in code loops
GLM-5.3-Flash is the first native multi-modal model in the GLM-5 series. Visual Coding (Visual Coding) doesn't just process images; it expands the boundaries of what programming can reach. For tasks such as front-end development, game development, and 3D simulation, the end product includes an interface, interaction, or virtual world that users can perceive. The visual ability is natively integrated into the model, enabling it to independently determine when to “observe” and use visual feedback to guide next steps.
We have developed a data synthesis pipeline for visual coding, focusing on improving the model's self-visual judgment and testing. The resulting trajectory requires the model to interact with the environment, examine its own output, and make iterative improvements. In terms of front-end coding, we also explore reinforcement learning based on environmental feedback, and further enhance GUI judgment capabilities through agent verification based on real user processes. This extends the scope of verification from functional correctness to rendering effects and interactive experience.
Surprisingly, Coding provided a model with a proxy to express and test its knowledge of the world. Building a coherent scene in Blender requires transforming knowledge of geometry, materials, lighting, and space into a 3D structure that is consistent from different perspectives. Without any external materials, the GLM-5.3-Flash operated independently for 16 hours to build a 400-square-meter professional chef's home and test kitchen.
In zCode, Browser Use Agent (BUA) and Computer Use Agent (CUA) further expand multi-modal capabilities, enabling models to work collaboratively between code, browser, and graphical interface while observing and verifying their own outputs—many issues are only revealed during rendering, interaction, or trial play. The following is a collection of experiments using GLM-5.3-Flash in zcode: from images to runnable projects, from two-hour movies to 8-minute films, and directly deliverable white-collar professional documentation.
A work partner that goes beyond coding
The performance of GLM-5.3-Flash in professional work is significantly improved compared to the same level model. For Office and document tasks such as PPTX, PDF, DOCX, and XLSX, new visual comprehension capabilities enable models to inspect and optimize their own output, and have stronger aesthetic judgment. The model can also compare its output with the visual context and expected results to achieve more effective self-verification and optimization—including more accurate assessment of presentation quality and aesthetic judgment.
We are specifically optimized for professional jobs in finance, law, etc. On the financial side, GLM-5.3-Flash can cover the complete process from source-based financial research and report generation to financial modeling and analysis, providing a traceable reference basis throughout the process. On the legal side, it is possible to review the cost, account and liability clauses in the contract and leave marks in accordance with lawyers' practice. It can also draft lawyers' letters, contracts, and litigation documents. The format and layout conform to practical specifications, and can be generated and delivered.
Running on domestic chips
In the past week, for the first time, we tried to use large domestic chip clusters to provide services in large-scale traffic. These chips are connected through self-developed high-bandwidth internet networks.
To overcome the problem of relatively limited computing power and memory capacity of a single chip, we built a dedicated inference engine based on SGLang. It is worth mentioning that the entire construction process was greatly accelerated by the Infra Agent driven by GLM-5.3. It helped engineers develop and optimize operators, diagnose performance bottlenecks, and improve deployment service stacks, forming a positive cycle of “model optimization system, system carrying model”.
The main bottlenecks of these chips are memory capacity and bandwidth. In particular, supporting up to 1M context lengths is very challenging. This requires us to carry out aggressive memory optimization for the underlying architecture, including custom optimization such as exchanging computing power for bandwidth and exchanging communication for video memory. Our technology stack combines techniques such as intra-node tensor parallelism for linear attention and LM head, ReplaySSM, W8A8 quantization, INT8/FP8/BF16 hybrid cache quantization, and layer split.
At the cluster level, we use a production-grade Encode-Prefill-Decode (EPD) split architecture to split multi-modal coding, prompt pre-filling, and token-by-token decoding into work pools that can be independently scheduled and independently expandable, thus achieving efficient and reliable services on domestic accelerators.
Compared with the initial baseline on the same hardware, end-to-end service performance has increased by 3 times, and hardware efficiency and single token cost have reached a level comparable to mainstream Nvidia GPUs. This proves that domestic chips can efficiently and economically support the reasoning requirements of cutting-edge models in large-scale scenarios.
Now, GLM-5.3-Flash is officially open sourced globally, connected to zCode and other coding platforms, and incorporated into the GLM Coding Plan (limited distribution of 10,000 experience cards per day) to simultaneously open API calls.