Python for Python

Marimo cloud/hosted offerings are separate from the OSS package — self-run is the free forever path. Vs classic Jupyter: different file format and reactivity model; some Jupyter extension habits do not transfer. Not a Spark/cluster notebook platform.

It is a library, not a trained model — you still bring weights and (for some connectors) a Roboflow API key. Roboflow ecosystem gravity is real if you lean on Inference. Vs OpenCV (low-level) and full ML platforms: Supervision is the high-level detection/annotation toolkit you import, not a training cloud.

Vs requests-cache / HTTPX built-ins: Hishel is the dedicated RFC caching layer with streaming-first design, not a framework and not a CDN. You still choose storage backends and correctness policy for authenticated or personalized responses.

NVIDIA stack gravity (TensorRT, Megatron, CUDA) — not an agent product. Serious runs need GPU/CUDA reality; CPU toy paths will not match the blog numbers. Ecosystem pull toward NVIDIA deploy targets even though the license is open.

Vs marimo: marimo is a reactive Python notebook runtime; Data Formulator is a dedicated AI viz/exploration app. Vs Pandas/Polars notebooks alone: you still bring model API keys for the agent path, and complex ETL still belongs in code. Microsoft research-to-product surface — expect UX churn; not a replacement for production BI (Metabase/Looker).

Not an agent product. Serious kernels need real GPU/NPU hardware and TVM-shaped mental models; CPU paths will not match CUDA blog numbers. Vs NVIDIA Model Optimizer: ModelOpt quantizes/prunes/distills HF models for TensorRT-LLM/vLLM; TileLang authors the kernels themselves. v0.1.x still drops legacy APIs — pin a release before production kernels.

The weights are licensed separately from the code. KittenTTS 2 uses the Stellon Labs Community License: free for research and non-commercial use, and for commercial use only until you or your affiliates pass USD 1M in annual revenue or total funding. The small 0.8 ONNX models are Apache-2.0. KittenTTS 2 wants about 6 GB of RAM on CPU or about 8 GB on a GPU, emotion control is beta, and only clone voices you have consent to use.

Releases are tagged beta and move fast, so pin versions in training scripts. Fine-tuning still needs a capable GPU for anything beyond small models, and multi-GPU support is newer than the single-GPU path. Watch the license split: the library is Apache-2.0, but the Studio UI is AGPL-3.0. The one-line installer pipes a remote script into your shell, so read it first.

A large share of the README is about bypassing bot detection and CAPTCHAs, which can break site terms — use it on sites you own or are allowed to automate. The API surface is huge, so pick one style and stick to it. For new JavaScript projects, Playwright is the more common default.

It is large and settings-heavy, so read the configuration docs closely. Default templates are plain and you will restyle them. Django-only by design. Development and some project resources have moved toward Codeberg and allauth.org.

Decorators are more verbose than Typer's type-hint style. Some behaviour changes between minor releases have broken downstream CLIs, so pin it. Rich output needs extra libraries.

uv tool install covers the same job and is much faster, so many people have switched. Each app gets its own environment, which uses more disk space. It is for apps, not libraries.

The keyboard-driven interface has a learning curve, and there is no mouse-first UI. Some formats need extra Python packages. It is terminal only (Windows through WSL).