[localai-org-maint-bot]

Description

Add Nanbeige4.3-33B to the model library and generated gallery with AtomicChat AD-Q4_K_M and Q8_0 builds for llama-cpp. The Q4 entry groups Q8 as a variant. Both use the embedded chat template, thinking enabled, a 32K default context, and the upstream chat sampling defaults. Document both install commands.

Sources: original model, GGUF builds. Neither master nor another open PR lists this model.

Notes for Reviewers

• Pin both downloads to revision b39515bab760dc8d6f7da7b4073847997971e2b7.
• Independently matched each HF API LFS SHA256 against the resolve response's x-linked-etag.
• Inspected both GGUF metadata and tensor tables: Llama architecture, 52 layers, supported tensor types, and embedded chat templates. The original config declares LlamaForCausalLM.
• AD-Q4_K_M SHA256: 409c92c4355c81928f0f2de5d49481fe7e4d0fa4d381aa28eddee0dc4b358c49.
• Q8_0 SHA256: 4106cbb9ba85da731f88b7493a90343e0d8cd6c77974ed1af00a97c5262c22bc.
• No full-weight download, local full-file hash, or inference run. No backend changes or old-model removals.

Validation passed:

go test ./scripts/models/...
go run ./scripts/models/cmd/modelcatalog --library gallery/models/index.yaml --out gallery/index.yaml
go test ./core/gallery -run 'Test(ModelLibrary|GeneratedGallery|CheckedInGallery|RealModelLibrary)' -count=1
go test ./core/gallery -run TestGallery -ginkgo.focus='gallery/index.yaml variant invariants' -count=1
git diff --check

All 2,294 entries round-trip; a second generation is byte-identical. All four variant invariants pass. Local tests reuse generated protobuf files from the prior cron worktree after confirming the proto sources are unchanged.

Signed commits

• Human DCO sign-off required; AI policy prohibits adding it on the submitter's behalf.
• Documentation updated under docs/content/.

[mudler-agent]

Landed on master in b669cda (batched gallery update). The per-model docs section was left out, since master no longer keeps those in model-gallery.md.