Update Basics tutorials (1xx) for model deprecations

[nealwu]

Note: branched from #732 to split up into smaller PRs.

Refreshes the Basics tutorials (101–104) off models slated for retirement on 2026-06-12, plus a few fixes found while running each notebook. First in a stack of five PRs splitting the tutorial refresh by section (Basics → Core Concepts → Cookbook Abstractions → Advanced → Deployment); this one is based on main.

Tutorials

• 101 Hello Tinker — switch to Qwen/Qwen3.5-9B-Base. A base model is the right fit for the raw tokenizer.encode → sample demo (a chat-tuned model drops into thinking mode on raw prompts). Tidy the sampling params.
• 102 First SFT — Qwen3-4B-Instruct-2507 → Qwen3.5-4B; Kimi-K2.5 → Kimi-K2.6 (with the kimi_k26_disable_thinking renderer); drop the deprecated name= arg from save_weights_and_get_sampling_client.
• 103 Async Patterns — align the narration with the asyncio.gather API the code actually uses (it previously described the synchronous sample()/.result() futures API).
• 104 First RL — meta-llama/Llama-3.1-8B → Qwen/Qwen3.5-9B-Base with the role_colon renderer (the recommended renderer for this base model).

Tooling

• Add a [tutorials] optional-dependencies extra (marimo, matplotlib) so a plain pip install tinker-cookbook stays lean; update tutorials/README.md to install via tinker-cookbook[tutorials].

Testing

Ran the notebooks end-to-end in marimo against the live Tinker service.

Note: diffs include some marimo serialization churn (the __generated_with version bump and trimmed cell-return tuples) from running the notebooks.

🤖 Generated with Claude Code