Skip to content

GuideAnts v0.9.19 — Qwen 3.8 27B

Latest
Latest

Choose a tag to compare

View all tags
@douglasware douglasware released this 14 Aug 23:09
v0.9.19
1e1d7b2
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

GuideAnts v0.9.19

Qwen 3.8 27B — curated install for Unsloth’s dense 27B vision model (262k context, hybrid thinking, in-GGUF draft-mtp), plus catalog-edit recovery when llama.cpp rejects a preset, clearer ReadWeb errors, and CWD-relative copy paths. AI / support images republished to GHCR for this cut.

Get started

  1. Install Docker Desktop (Windows/macOS) or Docker Engine 24+ with Compose (Linux). Windows needs WSL2.
  2. Download guideants-installer-v0.9.19.zip from this release.
  3. Unzip and run:
    • Windows: double-click guideants.cmd
    • Linux / macOS: chmod +x guideants.sh && ./guideants.sh
  4. Choose database layout, AI backend, and optional services when prompted.
  5. Open http://localhost:5107/ — first account becomes Admin.

Existing installs: relaunch the installer; if :main moved past your pins, accept the update prompt to pick up this channel. Volumes are kept.

Highlights

Qwen 3.8 27B

  • Qwen 3.8 27B (unsloth/Qwen3.8-27B-GGUF) in the curated llama catalog: vision mmproj, 262k native context, in-GGUF draft-mtp, and hybrid thinking with reasoning_effort (default xhigh)
  • Sampling defaults on the catalog row (temperature 1.0 / top_p 0.95 / top_k 20); Unsloth 4-bit typical 17–19GB RAM+VRAM (UD-Q4_K_XL)
  • Needs the republished AI images from this release (older llama.cpp crashes on draft-mtp / reasoning-preserve)

Local model catalog

  • Saving a router preset no longer 502s after llama-server rejects unknown keys; unrecognized options are stripped on respawn so the editor can recover
  • no-mmproj is rewritten to mmproj-auto=false

ReadWeb & notebook

  • ReadWeb failures report the actual status (401 / 403 / 404 / 429 / 5xx / timeout) instead of a generic error
  • GitHub hosts are not auto-excluded; excluded-host blocks are honored before fetch
  • Copy-path and attachment paths are CWD-relative so pasted paths work in chat tools

Images

  • AI / support images republished to GHCR (:main, :latest, and :v0.9.19) for this cut

Notes

  • Apple Silicon: prefer the slim AI backend (linux/amd64 under emulation).
  • Qwen 3.8 27B needs the republished AI images from this release (newer llama.cpp).
  • Operator details: docs/release-runbook.md, installer/README.md, deploy/azure/README.md.
Assets 3
Loading