---

base_model: unsloth/gemma-4-12b-it

pipeline_tag: text-generation

license: gemma

language:

• en

tags:

• gguf
• llama.cpp
• osint
• agent
• spotlight

---

Gemma-4 12B — Spotlight Orchestrator v4 (GGUF Q4_K_M)

Q4_K_M quantization (~7.4 GB) of the v4 Spotlight orchestrator — the servable

artifact for the 12B local tier of

Spotlight, an open-source OSINT

investigation harness. Merged from the

v4 LoRA adapter

(Unsloth-native merge with coherence check) and converted with llama.cpp.

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-v4

Hard-won on real investigation runs (48 GB Apple Silicon):

• --parallel 2 --ctx-size 81920 → 40,960 tokens per slot (a full

investigation loop reaches ~26k+). Bare -c 65536 without --parallel

allocates 4 full-size slots and OOMs.

• --no-cache-idle-slots — without it, the orchestrator's slot state is

saved/cleared on delegation and the restore fails mid-investigation

(failed to find available cells in kv cache). This bug masked model

capability across three tune iterations.

• --reasoning-budget 400 — matches the tune's reasoning cap; unbounded

reasoning consumes the whole token budget thinking and returns empty content.

• --jinja — required for tool calling; --cache-type-k/v q8_0 --flash-attn on

keep KV memory inside budget.

What it is trained to do

Drive the Spotlight orchestrator role: manage the phase pipeline (brief →

methodology → research → fact-check → report), delegate to

investigator/fact-checker subagents, present at human-approval gates with

non-empty syntheses, run the deterministic fact-check evidence validator before

the fact-check gate, and escalate bot-walled fetches to browser automation. See

the adapter card for the v4 dataset and the fixes over v3.

Status

Targeted checkpoint evaluation in progress (2026-07-10): resume-at-stall gate

presentation passed; fact-check replay and full chain-to-report pending.

Supersedes

buriedsignals/gemma4-12b-spotlight-orchestrator-v3-GGUF

once validation completes.