GLM-5.3
GLM-5.3 is Z.ai's flagship model for coding and agent work, built on the same base model as GLM-5.2 with all gains from post-training. It accepts text only, with a 1M-token context window and up to 128K output tokens, and always reasons.
Released
August 18, 2026
Type
llm
License
open-weight
GLM-5.3 License
Context
—
Z.ai states a “1M-token” context window.
Pricing
Input
$1.40 / 1M tokens
Output
$4.40 / 1M tokens
List price on the developer's own pages, checked . Cached input $0.26; cached input storage "Limited-time Free". No tiers shown.
Capabilities
Sources
Every figure on this page was checked against these pages on . A figure we could not source is left out.
- Z.ai docs: New Released (release notes) (opens in a new tab)
- Z.ai docs: GLM-5.3 (opens in a new tab)
- Z.ai docs: Pricing (opens in a new tab)
- Hugging Face: zai-org/GLM-5.3 model card (opens in a new tab)
- Z.ai (Zhipu AI): huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE (opens in a new tab)
Spotted something out of date? Tell us.
Links
In the news: GLM-5.3 and Z.ai (Zhipu AI)
MarkTechPost
Prime Intellect Launches Prime Inference: Serverless and Reserved Serving for Frontier Open Models
Prime Intellect has launched Prime Inference, a new serving platform designed for frontier open-source artificial intelligence models. This platform provides users with both serverless endpoints and reserved capacity, leveraging Prime Intellect's own Graphics Processing Units (GPUs) distributed across multiple datacenters. Prior to its public release, Prime Inference processed an internal volume of nearly one trillion tokens daily. This substantial internal traffic was generated from various applications including Reinforcement Learning (RL) rollouts, the creation of synthetic data, model evaluations, and the operation of long-running coding agents. Prime Inference functions as the serving layer within Prime Intellect's broader open training stack, which already includes post-training tools such as prime-rl, verifiers, and sandboxes. The introduction of serving capabilities completes this loop, enabling deployed models to generate production traces that can subsequently inform and improve the training process. Prime Intellect reports that its GLM-5.3 endpoint is among the fastest available on the OpenRouter platform, boasting a near-zero tool-call error rate and maintaining 100% uptime since its inception. The platform offers two primary serving modes: serverless endpoints are designed for handling variable and unpredictable demand, while reserved capacity is optimized for sustained and consistent workloads. Prime Inference is compatible with OpenAI's SDKs, allowing users to direct their OpenAI API calls to `https://api.pinference.ai/api/v1`. The system ensures high availability through automatic failover mechanisms that reroute traffic to healthy deployments across datacenters in the event of an issue. Future hardware support includes NVIDIA Blackwell GPUs, with Vera Rubin GPUs listed as coming soon. The platform features unified billing with team-level usage tracking, although per-model pricing details are still being finalized and published in the documentation. The underlying serving stack is a sophisticated integration of NVIDIA Dynamo, vLLM, Mooncake, and FlashInfer. This stack was developed in collaboration with Inferact and NVIDIA, with Prime Intellect actively contributing fixes back to the upstream projects. The primary target workload for Prime Inference is agentic AI, where a typical agent turn can add approximately 6,000 tokens to a prompt that already contains 140,000 tokens. Prime Intellect benchmarks this complex mix using the SemiAnalysis AgentX framework and simulates injected cold arrivals to test performance under load. A key architectural innovation is the disaggregation of prefill and decode operations, with each running on separate GPU groups. NVIDIA Dynamo manages the routing of tasks, while vLLM executes the models on their respective GPU groups. Decoders retrieve computed Key-Value (KV) data through NIXL. Prime Intellect claims this architecture results in nearly 40% lower p90 inter-token latency in their internal tests. Furthermore, the platform incorporates cache-aware routing, where Dynamo's KV-aware router dynamically weighs the overlap of cached KV prefixes against the volume of queued work to optimize performance and resource utilization.
MarkTechPost
Aikido Security Releases Altar-1: An Open-Weight Security Model Pruned From GLM-5.3 to 328 GB
Aikido Security has released Altar-1, its inaugural open-weight security model, on March 14, 2024. This model is a compressed iteration of Z.AI's GLM-5.3, engineered for deployment within customer-controlled infrastructure. Altar-1 serves as the foundational technology for Aikido Machine, the company's autonomous pentesting appliance, which is capable of operating on on-premise and air-gapped networks. The model's weights are publicly available on Hugging Face and can be executed using vLLM on a single node equipped with four NVIDIA H200 GPUs, indicating its deployability. The development addresses a critical challenge in the security sector: the inability for sensitive security context to remain within a company's network. Traditional closed-frontier models require data to be processed on external infrastructure, potentially exposing source code, architecture documentation, and unremediated security findings. This poses a significant issue for entities like banks with strict data-residency mandates or Operational Technology (OT) operators lacking internet connectivity. While open-weight models resolve data residency concerns, they often present deployment hurdles, particularly with Mixture-of-Experts (MoE) architectures where all experts must be stored, even if only a subset is actively used for a given task. Furthermore, security agents require substantial long-running context, and the associated KV cache competes with model weights for limited GPU memory. Altar-1 was constructed through a two-step compression process applied to GLM-5.3, a 753 billion parameter MoE model where each token routes to 8 out of 256 experts per layer, resulting in approximately 40 billion active parameters. The first step involved quantization, starting with the cyankiwi GLM-5.3-AWQ-INT4 checkpoint. This AWQ method stores routed expert weights in 4-bit precision, while activations are maintained at 16-bit. Shared components like attention, dense layers, and the output head are preserved in BF16 precision. The second step utilized Cerebras REAP (Router-weighted Expert Activation Pruning) for expert pruning. REAP evaluates experts based on router weights and output magnitude, not solely on their selection frequency. This process resulted in Altar-1 retaining 168 out of the original 256 routed experts per layer, effectively removing 88 experts, which constitutes a 34.4% reduction. Importantly, this pruning was achieved without any retraining of the model. The calibration for this process incorporated traces from Aikido's pentesting harness, alongside data from coding tasks, tool calling, reasoning exercises, and multilingual Wikipedia text. Aikido has stated that no customer data was utilized in this calibration phase. Each expert was scored based on its largest contribution to any single domain's routed workload, a methodology designed to safeguard the model's performance across diverse security-related tasks.
MarkTechPost
GLM-5.3-Flash vs Qwen3.8-Flash-Next: Two Chinese AI Labs Independently Converge on the Same Model Architecture
Two leading Chinese artificial intelligence labs, Z.ai and Alibaba's Qwen team, independently developed and released frontier open-weight models with strikingly similar architectures within a single week. Z.ai launched GLM-5.3-Flash, a 320 billion parameter multimodal Mixture-of-Experts (MoE) model featuring 18 billion active parameters. Concurrently, Alibaba's Qwen team unveiled Qwen3.8-Flash-Next, a 125 billion parameter model with 6 billion active parameters, which serves as a preview of the upcoming Qwen4 architecture. Despite their independent development, the configurations of these two models are nearly identical, indicating a potential convergence in cutting-edge AI model design. Both models employ a 3:1 hybrid attention mechanism, combining linear and full attention. They also utilize a compressed indexer to manage context, capped at 2048 tokens, and widen the residual stream into four gated branches. Furthermore, both models were trained using the Muon optimizer, with fused parameter matrices split prior to orthogonalization. GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series, released under the MIT license on Hugging Face. Z.ai tested it anonymously as Ox Alpha on OpenRouter, where it quickly became the most popular model of the week. This model was trained on a substantial 30 trillion token multimodal corpus and supports a context window of 1 million tokens. Z.ai claims that GLM-5.3-Flash outperforms its predecessor, GLM-5.2, across various benchmarks at one-tenth the cost, while achieving performance comparable to Claude Opus 4.8 on coding and agentic tasks. Its listed pricing is $0.15 per million input tokens and $0.50 per million output tokens. Qwen3.8-Flash-Next fulfills a similar role to Qwen3-Next in the Qwen3.5 release, acting as an early public preview of the next generation of Qwen architecture. The model card specifies a main model of 125 billion parameters, supplemented by an additional 51 billion n-gram embedding table, with 6 billion parameters activated per token. Its native context length is 262,144 tokens, which can be extended to 1 million tokens using YaRN. The Qwen team reported that the training of Qwen3.8-Flash-Next required approximately one-ninth of the compute resources needed for Qwen3.7-Plus. The technical report accompanying this release is titled “On the Design of Qwen3.8-Next Architecture: Evaluation, Efficiency, and Training Stability.” The shared architectural choices suggest a consensus among leading AI research labs on effective strategies for building powerful and efficient large language models, particularly in the realm of multimodal capabilities and context handling. The only notable point of disagreement among the described configurations is a single architectural detail, and one lab has reportedly dissented from the prevailing design choices.
MarkTechPost
Z.ai Releases GLM-5.3-Flash: A 320B-A18B Natively Multimodal MoE With a 1M-Token Context
Z.ai has released GLM-5.3-Flash, a new natively multimodal model that represents a significant advancement in the GLM-5 series. This model is characterized by its Mixture-of-Experts (MoE) architecture, featuring a total of 320 billion parameters with 18 billion active per token. A key feature is its exceptionally large context window, extending to 1,048,576 tokens, which allows for the processing of extensive amounts of information in a single input. GLM-5.3-Flash also boasts native multimodal capabilities, enabling it to process both image and video inputs. The model has been released under an MIT license, with its weights made available on Hugging Face, facilitating broader access and adoption by the research and development community. According to Z.ai's reports, GLM-5.3-Flash demonstrates superior performance compared to its predecessor, GLM-5.2, across various benchmarks and real-world applications. Notably, it achieves this enhanced capability at approximately one-tenth the price of previous models. In internal coding benchmarks, GLM-5.3-Flash performs competitively, achieving a score within half a point of Claude Opus 4.8. The model initially ran anonymously for its first week on platforms like OpenCode and OpenRouter, utilizing domestically produced Chinese AI chips for its operations. This release positions GLM-5.3-Flash as a cost-effective and powerful tool for a range of AI tasks. The deployment of GLM-5.3-Flash is structured across two primary tracks, ensuring accessibility for different user needs. The model's weights are publicly available on Hugging Face under the permissive MIT license, allowing organizations to self-host and integrate it into their infrastructure. Additionally, a priced and operational hosted API is already serving user requests, providing a more accessible entry point for those who may not have the resources for self-hosting. The default FP8 checkpoint requires approximately 306 GiB of storage before considering the KV cache, and the current vLLM implementation supports NVIDIA Hopper architecture and newer GPUs. This hardware requirement means that self-hosting is realistically within reach for mid-size to large organizations possessing at least an 8-GPU node or equivalent processing power, as well as AI-native startups that rent GPU capacity. For entities unable to self-host due to hardware or resource constraints, the API offers an economical solution where the focus shifts from hardware management to operational costs. Industries identified as having an immediate fit for GLM-5.3-Flash include software development and devtools, IT and Business Process Outsourcing (BPO) automation, financial services and insurance for document operations, enterprise business intelligence, back-office knowledge work, e-commerce, and any team responsible for shipping user interfaces at scale. Potential applications are diverse, ranging from developing repo-scale coding agents and terminal/browser-use agents to analyzing million-token logs and contracts, performing UI regression checks from screenshots, and enabling complex reasoning over spreadsheets, presentations, and dashboards that would typically require an Optical Character Recognition (OCR) to text pipeline.
The Hacker News
ThreatsDay: Gogs 10.0 RCE, n8n Workflow-to-RCE, $10M Reward, GLM-5.3 AI Exploit and More
The current threat landscape is characterized by vulnerabilities in trusted software and the increasing sophistication of attack methods, including those powered by artificial intelligence. A significant finding this week involves a Remote Code Execution (RCE) vulnerability in Gogs version 10.0, a popular self-hosted Git service. This flaw, stemming from a weak header check, allows attackers to execute arbitrary code on vulnerable servers, posing a substantial risk to organizations relying on Gogs for their code repositories. The ease with which this exploit can be leveraged lowers the barrier to entry for malicious actors seeking to compromise systems. Further compounding the security concerns is a critical workflow-to-RCE vulnerability discovered in n8n, an open-source workflow automation tool. This vulnerability allows for the execution of arbitrary code by manipulating n8n workflows, potentially leading to full system compromise. The nature of workflow automation tools means that a successful exploit could have cascading effects across integrated systems. The report also highlights the growing trend of legitimate applications and signed drivers being co-opted by malware to bypass defenses, making traditional security measures less effective. This tactic allows malicious code to blend in with normal system operations, evading detection. In parallel, the cybersecurity community is buzzing about a $10 million reward offered for the discovery of specific types of zero-day vulnerabilities. While the exact targets for this bounty are not detailed, such high-value rewards often indicate a focus on critical infrastructure, high-profile software, or advanced persistent threats (APTs). This initiative underscores the immense value placed on uncovering previously unknown exploits that could be used for espionage or disruption. Adding another layer of complexity, the report details the emergence of AI-assisted exploit research. Specifically, a vulnerability dubbed GLM-5.3 AI Exploit demonstrates how artificial intelligence can be used to discover and potentially weaponize security flaws. This development signals a new era in cybersecurity where AI is not only a tool for defense but also a powerful enabler for offense. The report also touches upon other security challenges, including exposed systems, the exploitation of old bugs, and unconventional hiding techniques employed by attackers, all contributing to a lower overall effort required to inflict damage. The confluence of these factors—vulnerabilities in widely used software, sophisticated evasion techniques, substantial exploit bounties, and AI-driven attack research—presents a challenging and evolving threat environment for individuals and organizations alike.