Introducing GPT-5.4 mini and nano: Lightweight inference variants of GPT-5.4 for coding, tool use, multimodal reasoning, and high-volume sub-agent workloads

Model Positioning and Core Optimization Goals

GPT-5.4 mini and GPT-5.4 nano are not independently trained architectures, but compact variants derived from the GPT-5.4 backbone via structured pruning, knowledge distillation, and operator-level kernel optimization — prioritizing low latency, high concurrency, and edge/on-device deployment feasibility.

Key Capabilities and Target Workloads

• Full support for Tool Calling v2 protocol, compatible with OpenAI Function Calling and JSON Schema-based tool schemas;
• Multimodal reasoning inherited from GPT-5.4, enabling joint image+text understanding (but no video stream input);
• On the HumanEval-Python benchmark, GPT-5.4 mini achieves 83.2% pass@1, while GPT-5.4 nano scores 76.5% pass@1;
• API throughput: GPT-5.4 nano delivers 128 req/s on A10 GPU (batch size=8, max_tokens=512), a 3.8× improvement over base GPT-5.4;
• VRAM footprint: GPT-5.4 nano requires only 4.2 GB VRAM in FP16 inference; GPT-5.4 mini uses 8.7 GB VRAM.

Deployment and Ecosystem Compatibility

• Available via OpenAI API under model IDs gpt-5.4-mini and gpt-5.4-nano, supporting streaming, logprobs, and parallel tool calling;
• Weights are not open-sourced; however, OpenAI released the inference SDK openai-inference-kit v1.4.2, featuring quantized loaders and a sub-agent orchestration middleware;
• Fully compatible with existing OpenAI infrastructure, including Assistants API, Threads API, and Code Interpreter plugin chains.