thinktecture/gemma3-4b-ft-nextera-q4_k_m

⚠️ Conference talk demo — not production weights.

This model accompanies a conference keynote on local on-device AI. Published as a reference for the fine-tuning patterns shown on stage — not a deployable artefact. No security audit, no SLA, pinned to the talk's state.

• Source repository: thinktecture-labs/local-multi-model-agent-slm
• Threat model + out-of-scope: SECURITY.md
• Licensing details: MODEL_LICENSES.md
• All five models in the stack: Collection — Local Multi-Model Agent — nextera fine-tunes

---

Gemma3-4B FT (Q4_K_M) — RAG Synthesis (+ Vision) — production

Base model	google/gemma-3-4b-it (4.3B params, multimodal: text + vision via mmproj)
License	Gemma Terms of Use
Provenance	llama-quantize from the F16 sibling GGUF (no separate training run — quantization only). See finetune/convert_gemma3_4b_to_gguf.sh.
File size	2.49 GB (vs 7.77 GB for F16) — ~3× memory-bandwidth headroom on decode
Hardware tested	RTX PRO 6000 (Blackwell sm_120), MBP M5 Max (Metal), DGX Spark (GB10 sm_121), Strix Halo (Vulkan/RDNA 3.5) — byte-deterministic across all four
Intended use	Production RAG response synthesis. Points-of-use: scenarios/<scenario>.json:synthesis_4b_gguf_ft. The vision channel uses the same GGUF (multimodal via the same mmproj as the F16 variant).
Out of scope	Tool calling (delegated to Qwen3.5-4B FT). Free-form chat without retrieved context.
Reference eval (Nextera, 2026-05-17)	Identical-quality to F16 on the 80-query RAG groundtruth set (MBP same-machine F16-vs-Q4_K_M A/B: 55/80 vs 54/80 = 1-query phrasing noise, zero semantic regression). Realized perf gains vs F16: RAG p50 -25% to -57%, image-query p50 -32% to -61% across the four-backend fleet.
Known failure modes	Same as F16 sibling. Q4_K_M-specific quantization artifacts not observed in our evals; would expect them most on rare-token tail behavior.