Skip to content

v2.0.0-beta.2: Qwen 3.5, WebLLM 0.2.86, pre-filled model source

Pre-release
Pre-release

Choose a tag to compare

View all tags
@pluginslab pluginslab released this 12 Oct 18:18
· 3 commits to v2 since this release
v2.0.0-beta.2
47504b5

Second beta of the v2 line, the full-featured GitHub release. The WordPress.org build stays on 0.11.x.

Pre-release. Not for use alongside the WordPress.org build (they share function names).

New

  • Model source is pre-filled. Settings → Model source now defaults to the MLC-AI addresses (Hugging Face for weights, GitHub for the compiled model libraries), so the local engine works without a setup step. The default applies until an administrator saves their own addresses. Clearing them still turns the local engine off.
  • Qwen 3.5 2B and Qwen 3.5 4B (experimental). Both are in the model dropdown, each in F16 and F32 builds. Qwen 3 1.7B stays the default.

Updated

  • WebLLM 0.2.82 → 0.2.86. Each model downloads a new compiled library (6 to 7 MB) once. Weights already in the browser cache are reused.

Fixed

  • Turning thinking off now works with Qwen 3.5. Qwen 3.5 ignores the /nothink prompt switch that Qwen 3 uses, so the local engine now also sends WebLLM's enable_thinking: false.
  • Model sizes in the dropdown were wrong. F16 and F32 builds are the same download and differ only in VRAM. The dropdown now shows both figures, taken from the model repositories and WebLLM.

Changed

  • The wait indicator says "Working..." instead of "Thinking..." and is left-aligned in a card, like an assistant message. "Thinking…" is now used only for the model's actual reasoning stream.

If you host the model files yourself

Compiled model libraries now live in <library URL>/v0_2_86/base/ and have new file names (for example Qwen3-1.7B-q4f16_1_cs1k-webgpu.wasm). Mirror that folder before upgrading.

Testing status

  • Unit tests, PHP lint and JS lint pass.
  • Qwen 3.5 2B scored 65/70 on the core ability tests through Ollama in a single run, the same as Qwen 3 1.7B (65/70 and 64/70 over two runs). The Ollama build is not the same quantisation as the in-browser one.
  • Qwen 3.5 4B has no reliable ability score yet. Its one test run was invalid, most likely because the test runner's way of turning thinking off does not work for Qwen 3.5.
  • The WebLLM upgrade and the Qwen 3.5 thinking fix have had only light manual testing in the browser.

Install

Upload agentic-admin.zip via Plugins → Add New → Upload.

Assets 3
Loading