# Qwen 3.8 in Rope Notes: Frontier Preview for Plan and Execute

> Source: <https://ropenotes.dev/blog/qwen-3-8-in-rope-notes/>
> Published: 2026-07-20 00:00:00+00:00

Three days after Moonshot’s Kimi K3, Alibaba’s Qwen team previewed **Qwen3.8-Max-Preview** at WAIC in Shanghai — a **2.4-trillion-parameter** multimodal flagship with a **1-million-token context window**. Alibaba positions it near the top of the frontier stack (their claim: second only to Anthropic’s Claude Fable 5). Independent leaderboards and a full model card are not public yet, so treat the ranking as the vendor’s until third-party numbers land.

If you already use Rope Notes’ agent for Dart, Flutter, or multi-file refactors, you can point a cloud provider at the preview today. Open weights are promised “soon,” with no date or license yet — so local 3.8 Max is not a thing yet.

| Fact | Status |
|---|---|
| Model ID | `qwen3.8-max-preview` |
| Parameters | 2.4T (Alibaba); activated-parameter count not published |
| Context | 1M tokens (per
|

There is **no new “Qwen 3.8B” dense twin** in this launch. The name is *Qwen 3.8* (family version), not an 8-billion-parameter sibling. For on-device work you still use the existing open-weight **Qwen3** line (`qwen3:8b`

and friends via Ollama / llama.cpp) until 3.8 Max weights appear.

Rope Notes is a local-first, IDE-style editor: rope-backed editing, Dart analysis, and an agent that can **Chat**, **Plan**, and **Execute** — with tools, then multi-file edit proposals you accept or reject. A long-context frontier model is useful for the Plan → Execute loop; a local Qwen3 keeps day-to-day work on-device.

| Qwen 3.8 Max strength | How Rope Notes uses it |
|---|---|
| Long-horizon coding + tools | Agent tool loop (`read_file` , search, definitions, …) over many turns |
| 1M context | Open tabs, `@` mentions, plans, and memory stay coherent longer |
| Multimodal input | Screenshot / layout feedback when your session includes image context |
| OpenAI-compatible API | Same Preferences → Agent flow as other cloud backends |
| Open weights later | Same UX once you host weights or a LAN OpenAI-compatible endpoint |

Alibaba markets strong coding and full-stack work relative to Qwen3.7-Max. Until independent scores exist, the practical test is your own Plan/Execute sessions — not launch-day leaderboard claims.

`qwen3.8-max-preview`

(not general pay-as-you-go at announce time).Open weights, when they land, change the local story. Until then, treat 3.8 Max as a cloud provider in Preferences.

You need a desktop or mobile build with **cloud agent backends** (web builds do not run the agent).

`https://dashscope-intl.aliyuncs.com/compatible-mode/v1`

`qwen3.8-max-preview`

If your Token Plan console shows a dedicated base URL (some Alibaba plans use a plan-specific host), use that URL instead — the model ID stays `qwen3.8-max-preview`

.

Provider configs live under `.rope_notes/preferences.toml`

. Share the *setup*, not the secret keys.

Rope Notes remains **Ollama-first** on desktop (llama.cpp is also first-class). For offline or source-sovereign sessions:

```
ollama pull qwen3:8b
# or a larger open Qwen3 tag your machine can hold
```

Then select the **Ollama (local)** or **llama.cpp (local)** provider and the pulled tag. That is the current Qwen3 open line — not Qwen3.8-Max-Preview.

`@lib/...`

or by dragging from Explorer.`plan-*.md`

/ checklist under `.rope_notes/`

.The 1M window helps across long checklists; still prefer targeted `@`

mentions so turns stay cheaper and sharper.

When Execute proposes changes, Rope Notes surfaces an inline diff review (Accept / Reject per file, or Accept all). The model does not write straight through your buffer — you still own the rope.

A common split:

`qwen3:8b`

(or similar)`qwen3.8-max-preview`

Switch providers from the agent panel; no project rewrite required.

On Android, cloud backends work even though local Ollama is a desktop feature. Use **session transfer** to move buffers and agent artifacts between machines, then keep the same Qwen Cloud provider selected.

Accurate expectations beat marketing:

`@`

mentions (with LSP diagnostics on those snippets)`/remember`

), optional cognitive memory hits, active skills`list_directory`

, `grep`

, `read_file`

, …) for on-demand deeper readsUse `@`

and Plan artifacts deliberately; a 1M window is not a reason to paste the whole repo by default.

Before long Execute sessions:

`.rope_notes/permissions.toml`

`.git`

and secret paths as neededFrontier models are most useful when they cannot touch everything.

`qwen3.8-max-preview`

`qwen3.8-max-preview`

)Qwen3.8-Max-Preview is available now as a hosted preview. In Rope Notes it is a Preferences change — and a natural cloud counterpart to the local Qwen3 models you may already run on Ollama.

*Rope Notes — a high-performance, local-first editor with a real AI agent, Dart analysis, and optional P2P sync. ropenotes.dev*