capture_date_utc stringdate 2026-08-26 00:00:00 2026-08-26 00:00:00 | method_version stringclasses 1
value | corpus_version stringclasses 1
value | gpu stringclasses 1
value | driver float64 611 611 | llamacpp_build stringclasses 1
value | model stringclasses 5
values | gguf_file stringclasses 5
values | device stringclasses 1
value | ngl int64 99 99 | ctx int64 8.19k 8.19k | server_load_s float64 2.58 26 | prompt_tokens int64 1.49k 2.46k | cold_total_ms float64 284 1.08k | warm_total_ms_median float64 126 585 | warm_reps int64 5 5 | cold_prefill_tok_s float64 2.94k 15.5k | decode_tok_s float64 86.3 287 | accuracy_pct float64 70.8 95.8 | accuracy_n int64 24 24 | vram_before_mib int64 864 1.51k | vram_after_mib int64 4.03k 13.1k |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
2026-08-26 | ha-voice-brain-v1 | 1.0.1 | NVIDIA GeForce RTX 5080 | 610.88 | build 10507, commit 95c409c13 | gpt-oss-20b-mxfp4 | gpt-oss-20b-MXFP4.gguf | gpu-rtx5080 | 99 | 8,192 | 19.61 | 1,494 | 1,081.8 | 585.4 | 5 | 2,942.4 | 211.7 | 91.7 | 24 | 1,511 | 13,112 |
2026-08-26 | ha-voice-brain-v1 | 1.0.1 | NVIDIA GeForce RTX 5080 | 610.88 | build 10507, commit 95c409c13 | qwen2.5-14b-q4km | qwen2.5-14b-instruct-q4km-ollama-blob.gguf | gpu-rtx5080 | 99 | 8,192 | 26.02 | 2,106 | 1,018.4 | 371 | 5 | 3,843.1 | 86.3 | 95.8 | 24 | 1,495 | 11,571 |
2026-08-26 | ha-voice-brain-v1 | 1.0.1 | NVIDIA GeForce RTX 5080 | 610.88 | build 10507, commit 95c409c13 | llama3.2-3b-q4km | llama-3.2-3b-instruct-q4km-ollama-blob.gguf | gpu-rtx5080 | 99 | 8,192 | 2.83 | 2,459 | 283.9 | 125.9 | 5 | 15,469.2 | 286.6 | 70.8 | 24 | 864 | 4,032 |
2026-08-26 | ha-voice-brain-v1 | 1.0.1 | NVIDIA GeForce RTX 5080 | 610.88 | build 10507, commit 95c409c13 | qwen3-4b-instruct-2507-q4km | Qwen3-4B-Instruct-2507-Q4_K_M.gguf | gpu-rtx5080 | 99 | 8,192 | 2.58 | 2,104 | 325.9 | 143.6 | 5 | 12,352.1 | 232.5 | 83.3 | 24 | 868 | 4,741 |
2026-08-26 | ha-voice-brain-v1 | 1.0.1 | NVIDIA GeForce RTX 5080 | 610.88 | build 10507, commit 95c409c13 | qwen2.5-7b-q4km | sha256-2bada8a7450677000f678be90653b85d364de7db25eb5ea54136ada5f3933730 | gpu-rtx5080 | 99 | 8,192 | 4.86 | 2,106 | 556.8 | 291.1 | 5 | 8,243.6 | 127.5 | 87.5 | 24 | 864 | 5,897 |
Home Assistant Local LLM Voice Benchmark
Per-model tool-call accuracy and component latency for running a Home Assistant voice assistant against local LLMs.
Measured per-model tool-call accuracy and component latency for running a Home Assistant voice assistant against local LLMs.
Broken out by pipeline component rather than reported as one opaque round trip, so you can tell whether your latency is wake-word, speech-to-text, the model, or text-to-speech before you go optimizing the wrong stage.
Source and license
Canonical page, full method and change log: https://techfuelhq.com/tutorials/home-assistant-local-llm-voice-2026/
Licensed CC BY 4.0. Attribution: "TechFuelHQ" linking to the canonical page.
Corrections welcome. If a row is wrong it gets fixed against the primary source and the correction is logged.
- Downloads last month
- 30