What's New

AWS has added two new foundation models — GLM-5.1-FP8 from Z.ai and Phi-4-mini-instruct from Microsoft — to Amazon SageMaker JumpStart, making them available for enterprise deployment on AWS infrastructure. GLM-5.1-FP8 targets agentic software engineering workflows requiring deep, iterative reasoning, while Phi-4-mini-instruct delivers capable reasoning and multilingual support in a compact, resource-efficient form factor. Both models can be deployed with minimal friction through SageMaker Studio or the SageMaker Python SDK.

How It Works

• GLM-5.1-FP8 (Z.ai): Uses FP8 quantization to reduce memory footprint while preserving model quality, enabling efficient deployment of a large agentic coding model on AWS GPU instances.
• GLM-5.1-FP8 agentic loop: Designed for sustained multi-round optimization, the model can iterate hundreds of reasoning cycles over a single task — such as repository-level code generation or complex debugging — improving its output with each pass.
• GLM-5.1-FP8 task scope: Handles terminal-level tasks, file system interactions, and code review pipelines, making it suitable for AI-powered development environments that require long-horizon problem-solving.
• Phi-4-mini-instruct (Microsoft): A compact small language model (SLM) optimized for strong reasoning, mathematics, and logical inference while operating within tight memory and latency budgets.
• Phi-4-mini-instruct multilingual support: Supports 24 languages and native function calling, enabling structured tool use and multilingual conversational applications in a single deployment.
• SageMaker JumpStart deployment: Both models are accessible via the Models section of SageMaker Studio or programmatically through the SageMaker Python SDK, allowing one-click or code-driven deployment to a managed endpoint in the customer's AWS account.

Why It's Important

• Agentic AI at scale: GLM-5.1-FP8 addresses the growing enterprise need for AI systems that can autonomously execute multi-step software engineering tasks, reducing developer toil in code review, debugging, and repository management.
• Efficiency without sacrifice: Phi-4-mini-instruct demonstrates that strong reasoning and multilingual capability no longer require large, expensive models, lowering the cost barrier for deploying capable AI in constrained environments.
• Broader model choice on AWS: Adding models from Z.ai and Microsoft expands the diversity of providers in SageMaker JumpStart, giving customers more options to match model strengths to specific workload requirements without leaving the AWS ecosystem.
• Edge and latency-sensitive use cases unlocked: Phi-4-mini-instruct's compact footprint makes it viable for near-edge deployments and real-time applications where response latency is a hard constraint.
• Reduced operational complexity: SageMaker JumpStart abstracts infrastructure provisioning, allowing teams to focus on application logic rather than model serving infrastructure.

How It's Different

• GLM-5.1-FP8 vs. general-purpose coding models: Unlike standard code-completion models, GLM-5.1-FP8 is explicitly designed for iterative, agentic loops — it improves solutions over hundreds of rounds rather than producing a single-shot response.
• FP8 precision advantage: The FP8 quantization of GLM-5.1 reduces GPU memory requirements compared to FP16/BF16 variants, enabling deployment on fewer or smaller instances while maintaining competitive accuracy.
• Phi-4-mini-instruct vs. larger reasoning models: Phi-4-mini-instruct achieves strong reasoning and math benchmarks in a fraction of the parameter count of models like GPT-4 or Claude, making it uniquely suited for cost- and latency-sensitive deployments.
• Native function calling in a small model: Many compact SLMs lack reliable function-calling support; Phi-4-mini-instruct includes it natively, enabling structured agentic tool use without a large model.
• Multi-provider diversity in JumpStart: These additions bring models from Z.ai (a less commonly represented provider on AWS) and Microsoft alongside existing offerings from Anthropic, Meta, and others, broadening architectural and licensing diversity.

When to Prefer It

• Choose GLM-5.1-FP8 when building automated software engineering agents that need to navigate codebases, execute terminal commands, and iteratively refine solutions over many reasoning steps.
• Choose GLM-5.1-FP8 for AI-assisted code review pipelines where the model must analyze pull requests, suggest fixes, and re-evaluate changes across multiple rounds of feedback.
• Choose Phi-4-mini-instruct when deploying AI in memory-constrained environments such as smaller GPU instances, edge servers, or cost-optimized inference endpoints where a large model is impractical.
• Choose Phi-4-mini-instruct for multilingual customer-facing chatbots or support applications that require reasoning and math capabilities across up to 24 languages with low latency.
• Choose Phi-4-mini-instruct when building tool-calling agents that need structured function invocation but must operate within strict latency SLAs or tight infrastructure budgets.
• Use either model via SageMaker JumpStart when your team wants to avoid custom model-serving infrastructure and prefers a managed, click-to-deploy experience integrated with existing AWS IAM, VPC, and monitoring tooling.

Availability

• General Availability: Both GLM-5.1-FP8 and Phi-4-mini-instruct are generally available in Amazon SageMaker JumpStart as of May 14, 2026.
• Access method: Models are accessible through the Models section of Amazon SageMaker Studio or via the SageMaker Python SDK for programmatic deployment.
• Regional availability: Specific supported AWS regions are not detailed in the announcement; customers should consult the SageMaker JumpStart documentation or the AWS Regional Services table for the current region list.
• Pricing: Costs follow standard SageMaker real-time inference pricing based on the instance type selected for the endpoint; no separate model licensing fee is mentioned, but customers should verify any third-party model terms with Z.ai and Microsoft.
• Limitations: No explicit context window sizes, maximum token limits, or fine-tuning support details are provided in the announcement; refer to model-specific documentation for these constraints.