tomvaillant/gemma4-12b-spotlight-orchestrator-v5-GGUF

Gemma-4 12B — Spotlight Orchestrator v5 (GGUF Q4KM)

Q4KM quantization of the v5 Spotlight orchestrator — the servable artifact for the 12B local tier of Spotlight. It is merged from the v5 LoRA adapter with Unsloth's native merge path, coherence-checked, and converted with llama.cpp.

Files

• —gemma-4-12b-spotlight-orchestrator-Q4_K_M.gguf — the single 7.381 GB, text-only Q4KM deployment artifact. Vision weights are intentionally not included.

Serving — the flags are load-bearing

llama-server -m gemma-4-12b-spotlight-orchestrator-Q4_K_M.gguf \
  --port 8093 --host 127.0.0.1 \
  --ctx-size 81920 --parallel 2 --no-cache-idle-slots \
  --n-gpu-layers 999 --jinja --flash-attn on \
  --cache-type-k q8_0 --cache-type-v q8_0 \
  --reasoning-budget 400 \
  --alias gemma-4-12b-spotlight-v5

These settings are part of the tested configuration:

• —--parallel 2 --ctx-size 81920 gives each slot 40,960 tokens, matching the tuning context and leaving enough room for a full investigation loop.
• —--no-cache-idle-slots prevents slot save/restore failures during subagent delegation.
• —--reasoning-budget 400 prevents reasoning from consuming the response budget before a tool call or gate presentation is emitted.
• —--jinja enables tool calling; q8 KV and flash attention keep memory use inside the local 48 GB target.

What it is trained to do

Drive Spotlight's brief → methodology → research → fact-check → report pipeline; delegate constrained investigation and fact-check work; obey human gates; and recover cleanly when deterministic validators reject structured inputs. The v5 dataset contains the report-execution recovery contexts that v4 fumbled, but the production harness no longer asks any model to edit the 20 KB HTML template: Spotlight renders all three report artifacts deterministically from validated findings.json and fact-check.json.

Validation status

Released 2026-07-10 after the preserved v4 failure fixture passed the canonical report finalizer: evidence-chain validation, deterministic Markdown/HTML/evidence map rendering, report-gate validation, and byte-identical repeat rendering. The same finalizer also passed the completed 26B case. A fresh full 12B E2E is the post-release validation; report file construction is no longer part of that model-capability bet.

Install

curl -fsSL https://spotlight.buriedsignals.com/install-spotlight.sh | bash

Direct GGUF download:

hf download tomvaillant/gemma4-12b-spotlight-orchestrator-v5-GGUF \
  gemma-4-12b-spotlight-orchestrator-Q4_K_M.gguf

Limitations

The model does not replace source verification. Spotlight's deterministic fact-check validator and report finalizer are required deployment components. Tool use is tuned against Spotlight's harness schema and may not transfer to other agent frameworks without adaptation. The Q4KM artifact is text-only; use a separate vision-capable fact-check tier for scanned documents or imagery.