GPT-5.5 vs Kimi, DeepSeek and Mistral: model swap tutorial

Pamela Fox compares GPT-5.5, Kimi K2.6, DeepSeek V4 Flash and Mistral Large 3 through Python examples on Microsoft Foundry. The workshop tests application behavior rather than relying on benchmark rankings. Setup uses Python, uv and a .env file, with GitHub Codespaces as another option.

The examples cover letter counting, RAG groundedness and image interpretation before moving into tool calls. In Fox's demonstrations, a revised prompt fixes Mistral's counting answer, but stronger source instructions still leave unsupported information in its RAG response. DeepSeek V4 Flash skips the image tests because it accepts text only. These results describe the workshop's tasks, rather than a universal ranking.

A LangChain v1 travel planning AI agent exposes differences in execution: GPT-5.5 requests flight and hotel searches in parallel, while other runs take extra turns or omit required steps. Fox then demonstrates string checks, LLM judges and the need to calibrate them against human labels, and ASSERT test generation from a YAML specification. DSPy produces a prompt for Mistral using explicit output constraints and evaluation checks.

For local LLM testing, Fox says the examples can point to an Ollama server with different model names. She expects general API compatibility; the main demonstrations use hosted deployments.