Small, runnable examples from RUNTIME on YouTube, with setup instructions, saved outputs and honest limits.
New to this? Start here. Already using Python? Open the technical index.
| Guide | What you build | Start / watch |
|---|---|---|
| xLLM-Loop | Compare full and shared attention caches, inspect exact outputs, and recompute saved memory and timing results | Code, measurements and setup |
| Ornith 1.5 9B vs Qwen | Replay three saved coding tasks, inspect browser failures and distinguish working code from unfinished runs | Tasks, artifacts and local-run notes |
| MiniCPM5 1B vs 2B | Recompute twelve saved text checks and inspect an executed native-tool workflow | Results, sources and local-run notes |
| CLM / Jev / Laya | Inspect saved decisions, explore the offline visualizer, and run deterministic simulation experiments | Research, code and viewer |
| Tev1 vs Qwen | Recompute 230 saved decisions, inspect failures, and follow an optional local rerun recipe | Code, evidence and setup · Video |
| Kev vs Laya | Inspect 180 recorded decisions, compare architectures and resources, recompute findings | Research and analysis |
| Laya | Classify requests, inspect uncertainty, customize categories, optionally draft with a local LLM | Guide · Video |
| Needle 3 | Convert seven synthetic requests into proposed tool calls and inspect failures | Guide · Video |
For the advanced experiment from our Dream-RSI episode, use the separate dream-rsi-experiment repository.
Laya and Needle were checked on Python 3.12 / Linux ARM64 / CPU. Tev1 measurements used a local NVIDIA GB10 GPU; its exported saved-result analysis is checked offline on Python 3.12, while the optional patched inference recipe still needs fresh qualification. See each guide for its exact environment. Other platforms are not qualified. Model weights download separately.
These small synthetic demonstrations are not performance or accuracy benchmarks. Nothing sends messages or changes accounts.
Use Code → Download ZIP above and extract the folder, or clone:
git clone https://github.com/Runtime-weekly/runtime-tutorials.git
cd runtime-tutorialsThen open one guide and create its virtual environment inside that example folder. The repository root has no shared install command. No RUNTIME account, API key or private service is required; Laya's optional drafting step needs your own local language-model server.
The Laya tutorial ZIP contains the video examples with updated, tested dependency pins. The Laya and Needle tutorial library includes those two guides, historical outputs and navigation. The Tev1 vs Qwen and Kev vs Laya research companions are available in the current repository ZIP or clone. Older releases are retained for reference; use these current downloads for a new installation.
| Folder / page | Contents |
|---|---|
| minicpm5/ | Frozen synthetic tasks, final answers, native tool traces and qualified reproduction notes |
| START_HERE.md | Downloading files, opening a terminal and creating an environment |
| tev1/ | Exact synthetic cases, model revisions, responses, timing, offline verification and optional local inference |
| kev-laya/ | Frozen synthetic cases, saved decisions, architecture and resource measurements, standard-library analysis |
| laya/ | Laya code, requirements, instructions and saved results |
| needle/ | Needle code, test sentences, instructions and saved results |
| docs/ | Glossary, troubleshooting and direct technical links |
The Laya example code is unchanged from the video; dependency pins were updated after an advisory review. The CPU examples were rerun in a fresh environment and their labels, rounded probabilities and gate decisions matched. The optional LLM integration was not rerun with the updated packages. The Needle export was installed in a fresh Python environment and all seven requests were rerun; proposed calls and warnings matched the recorded demonstration. Package downloads used existing caches. This does not qualify a cold download or other operating systems. See each guide for source/version details and actual failures.
Original RUNTIME example code and accompanying tutorial documentation are available under the MIT License. Upstream code, model weights and dependencies retain their own licenses. This license does not cover channel videos, narration, portraits, logos or other brand assets. Models and environments are downloaded separately; they are not bundled here.
Roundups and newsletters are kept in a separate archive. For a problem with this tutorial, open an issue with the command, platform and a small synthetic example. Remove secrets, personal details, private paths and customer data before sharing logs.