Ollama Cloud Rolls Out Z.ai GLM-5.3 and GLM-5.3 Flash with Private Hosting

GLM-5.3 is a frontier model developed by Z.ai that maintains the same base architecture as GLM-5.2 while achieving substantial performance increases exclusively through post-training methods.

Ollama has completed the full rollout of both GLM-5.3 and its efficient multimodal sibling GLM-5.3 Flash on the company's cloud platform. These models are now listed as glm-5.3:cloud and glm-5.3-flash:cloud respectively. The deployment emphasizes private hosting options located in the United States and Europe. A strict zero data retention policy applies to all interactions with these models on the platform. This setup allows developers to utilize the models without concerns over data storage or retention by the provider.

Background on Z.ai and Previous GLM Releases

Z.ai has positioned the GLM series as a core component of its frontier research program focused on coding and agent capabilities. The GLM-5.2 release established foundational infrastructure that included training stacks and inference optimizations. Subsequent work shifted emphasis toward refining outputs after the initial pre-training phase. This incremental strategy enables targeted capability lifts while preserving the core model weights.

Company documentation notes that the GLM-5.2 effort concentrated on building the overall technical stack. The current release therefore isolates post-training as the variable under study. Such an approach reduces computational overhead associated with full pre-training cycles and allows faster iteration on alignment and capability scaling.

Announcement of the Ollama Cloud Integration

The official announcement from Ollama highlighted the private and fast hosting available in US and Europe data centers. The zero data retention policy aligns with the provider's standard privacy commitments for all cloud-hosted models. Developers received immediate access through familiar tooling without additional configuration steps.

The rollout announcement specified that the models follow Ollama’s existing privacy policy without exception. Geographic hosting options ensure compliance with regional data residency requirements. Latency remains low for users in the supported regions due to the distributed infrastructure.

Technical Details of GLM-5.3

GLM-5.3 contains 753 billion total parameters. All measured gains relative to GLM-5.2 derive from post-training refinements applied to the unchanged base model. This design choice isolates the contribution of alignment and fine-tuning stages.

The post-training process incorporates additional scaling of existing techniques developed during the GLM-5.2 cycle. Resulting improvements appear across coding benchmarks and terminal interaction tasks. No modifications to the underlying pre-trained weights occurred between the two versions.

Specifications of the GLM-5.3 Flash Variant

GLM-5.3 Flash operates with 321 billion total parameters in one reported configuration and 320 billion in another, alongside 18 billion active parameters during inference. The model constitutes Z.ai’s first natively multimodal release capable of processing text, image, and video inputs within a single architecture.

Before the public launch, Z.ai evaluated the Flash variant under the anonymous identifier Ox Alpha on third-party platforms including OpenRouter and OpenCode. This testing phase allowed collection of usage signals without revealing the originating organization. The multimodal support differentiates the Flash model from the text-only GLM-5.3 flagship.

Comparison of key specifications between GLM-5.3 and GLM-5.3 Flash

Model	Total Parameters	Active Parameters	Multimodal Support	Primary Strength
GLM-5.3	753B	Not specified	Text only	Coding and terminal tasks
GLM-5.3 Flash	321B	18B	Text, image, video	Efficient multimodal inference

Performance Improvements Reported

Z.ai reported a 50 percent improvement on its internal Code Bench evaluation compared with GLM-5.2 results. The company attributes the entire delta to refinements introduced during the post-training stage. This benchmark focuses on code generation and understanding tasks relevant to software engineering workflows.

The Terminal-Bench increase from 4.6 to 28.3 demonstrates measurable progress in handling complex command-line and scripting scenarios. Such gains suggest stronger emergent abilities in agent-like interactions. The documented scores appear directly on the model library pages maintained by Ollama.

Access and Usage Options

Developers interact with the models through multiple supported interfaces. Command-line invocation provides the simplest entry point for individual experimentation. API integration supports production deployments and orchestration with external tools.

1. Invoke the model via the command ollama run glm-5.3:cloud or the corresponding Flash tag
2. Initiate agent sessions through the ollama launch interface for structured workflows
3. Authenticate requests with API keys when embedding the models in custom applications
4. Confirm regional hosting and retention settings through the provider documentation prior to deployment

Implications for Enterprise and Developer Users

Private hosting combined with zero data retention addresses compliance requirements common in regulated industries. Enterprises can evaluate frontier capabilities without transmitting proprietary data to external retention systems. The geographic options further support adherence to data localization mandates.

Multimodal input handling in the Flash variant extends applicability to document analysis, video summarization, and mixed-media agent tasks. Developers can prototype applications that combine visual reasoning with code generation within a single model call. This reduces the need for separate specialized models in certain pipelines.

Reactions and Official Statements

GLM 5.3 and GLM 5.3 Flash (previously Ox Alpha) are fully rolled out on Ollama's cloud. Private. Fast. US and Europe hosted. Zero data retention.Ollama

The statement from the Ollama account underscores the emphasis on privacy and geographic hosting options. Z.ai’s accompanying blog post focuses on the post-training methodology as the sole source of capability advancement. Both communications avoid broader claims about future scaling trajectories.

Outlook for Subsequent Developments

Continued emphasis on post-training may produce additional releases that build on the current base without full pre-training cycles. Platform availability through providers such as Ollama could generate usage data that informs subsequent alignment choices. Observers will track whether the multimodal Flash variant achieves parity with specialized vision-language models on dedicated benchmarks.

Integration patterns for agent frameworks and retrieval-augmented generation systems represent areas for near-term experimentation. The zero-retention hosting model may influence similar offerings from competing cloud providers seeking to attract privacy-conscious users.