---

base_model: Qwen3-8B

tags:

• qwen3
• home-assistant
• gguf
• ollama
• structured-output
• ha-action-v3

library_name: gguf

---

Home Assistant Specialist v0.4 — Qwen3-8B Q6_K

GGUF artifact for local Ollama inference.

File: Qwen3-8B.Q6_K.gguf

Quantization: Q6_K

Task: strict ha-action-v3 Home Assistant JSON planning.

Dataset provenance: V4 combines validated rows from acon96/Home-Assistant-Requests, acon96/Home-Assistant-Requests-V2, and project-curated contract cases. See the V4 dataset card for counts and filtering details.

Benchmark

Quality matched Q5 but incurred CPU offload and higher latency.

| Metric | Result |

|---|---:|

| Exact match | 97.36% |

| Safety OK | 98.32% |

| Entity not invented | 100.00% |

| Mean latency | 2.348 s |

The evaluation uses the fixed 417-case test split from tuxevil/Home-Assistant-Requests-V4.

Same-quantization model comparison

| Quantization | Qwen3-8B exact | Qwen3-4B exact | Delta | Qwen3-8B safety | Qwen3-4B safety | Delta |

|---|---:|---:|---:|---:|---:|---:|

| Q4_K_M | 91.85% | 87.53% | +4.32 pp | 93.05% | 88.49% | +4.56 pp |

| Q5_K_M | 97.36% | 95.68% | +1.68 pp | 98.32% | 96.64% | +1.68 pp |

| Q6_K | 97.36% | 92.81% | +4.55 pp | 98.32% | 94.48% | +3.84 pp |

| Q8_0 | Not published | 92.57% | — | Not published | 93.76% | — |

Hugging Face / Transformers reference

| Model | Base 4-bit exact | Adapter 4-bit exact | Merged FP16 exact | Base safety | Adapter safety | Merged FP16 safety |

|---|---:|---:|---:|---:|---:|---:|

| Qwen3-8B | 0.00% | 99.28% | 99.20% | 84.17% | 100.00% | 99.50% |

| Qwen3-4B | 0.00% | 98.08% | 88.25% | 75.54% | 98.80% | 89.69% |

* Qwen3-8B FP16 published reference retains exact-match and safety metrics; other FP16 fields were not retained in the artifact manifest.

Quantization comparison

| Quantization | Exact match | Safety OK | Mean latency | Deployment note |

|---|---:|---:|---:|---|

| Q4_K_M | 91.85% | 93.05% | 1.536 s | Full GPU |

| Q5_K_M | 97.36% | 98.32% | 1.580 s | Full GPU; preferred |

| Q6_K | 97.36% | 98.32% | 2.348 s | CPU offload; slower |

Ollama

The repository includes a Modelfile. Example:

ollama create home-assistant-specialist-v0.4-8b-q6 -f Modelfile

Use the exact model tag and verify the runtime with the full held-out benchmark before production use.

Safety

Do not execute generated JSON directly. Enforce status, service, entity, and parameter allowlists against live Home Assistant context.