Microsoft debuts Phi‑4 small language models on Azure AI with lower per token pricing and enterprise controls

Microsoft launched two Phi‑4 small language models on Tuesday for Azure AI, adding 4 billion and 14 billion parameter variants that target low latency inference and tight cost envelopes. The models are available immediately in the Azure AI Model Catalog across the United States and European Union regions, with list prices starting at 0.05 dollars per million input tokens and 0.20 dollars per million output tokens for the 4B model, according to Microsoft.

The company said the 14B model is priced at 0.20 dollars per million input tokens and 0.80 dollars per million output tokens. Microsoft is pitching the release as an enterprise friendly alternative to larger closed models for high volume chat, classification, and tool use.

Pricing and benchmarks

Microsoft said Phi‑4 14B scores 71 on MMLU and 74 percent on GSM8K with few shot prompts. The company said this outperforms Meta Llama 3 8B on the same settings, which Meta reported in April 2024 at about the high sixties on MMLU and the high sixties on GSM8K. Microsoft also cited latency under 140 milliseconds at 8 thousand token output on a single NVIDIA L4 in Azure, and under 60 milliseconds on an H100 with tensor parallelism enabled, measured in its internal tests.

Both models ship with structured output support, function calling, and JSON schema validation in the Azure inference API. Microsoft said audit logs route to Microsoft Purview by default, and that customers can enable zero data retention with customer managed keys through Azure Key Vault.

Why this matters now

Small models have become a cost and latency lever for production workloads that do not require the breadth of frontier systems. Gartner said in a May 2026 note that enterprises are shifting 35 percent of new generative AI tasks to models under 20 billion parameters, up from 7 percent in 2024, primarily for privacy and throughput reasons.

Microsoft’s move widens its catalog beyond GPT models from OpenAI, which remain premium priced for many scenarios. It also answers demand from regulated buyers in the European Union as obligations under the EU AI Act begin phasing in during 2026, where providers and deployers must meet transparency and risk management requirements.

Competitors are following a similar script. Meta released Llama 3 in April 2024 as open weights, and Mistral has pushed smaller instruction tuned models through its API and Azure partnership. Cohere and Google have also promoted compact models for classification and retrieval tasks, with token based pricing that targets high volume enterprise usage.

If adopted, cheaper small model tiers on hyperscale clouds could shift budget from inference to data and integration work, which favors vendors with analytics, governance, and vector infrastructure already in place. That dynamic benefits platforms like Azure, Google Cloud, AWS, and Databricks that bundle model endpoints with storage, observability, and security controls.