Did you keep up this week?

Week of 19 Oct 2026. 5 questions on this week's AI news, plus one tool worth trying.

How many new partner integrations did Figma add to its MCP server?
Who at Figma said MCP is not a perfect protocol?
What does Claude Code Review do?
What did Claude Security add in its latest update?
What kind of data does Notified's new MCP server expose?

0 of 5 answered

Tool of the week: vLLM v0.32.0

vLLM is an Apache-2.0 engine for running and serving open LLMs on your own GPUs with high throughput and an OpenAI-style API. The v0.32.0 release (Oct 18) adds a structured-decision endpoint, 4-bit KV cache on more GPUs, faster cold starts via precompiled kernels, and new models such as EmbeddingGemma 2.

If you self-host open models like DeepSeek V4.1 Flash or Qwen3.8, this release makes serving cheaper on memory and quicker to restart.

Try it in 5 minutes

Use a machine with a supported GPU and a Python version newer than 3.10, since this release dropped 3.10.
Install vLLM with uv as shown in the README.
Run vllm download-kernels to fetch precompiled kernels so startup is faster.
Start the server with any small open model from Hugging Face (see the last command).
Point any OpenAI SDK client at the local server and send a test chat request.

bash

uv pip install vllm
vllm download-kernels
# OpenAI-compatible API on localhost:8000
vllm serve <hugging-face-model-id>

Open vLLM v0.32.0