| --- |
| license: other |
| library_name: gguf |
| tags: |
| - gguf |
| - qwen |
| - eliza-1 |
| - local-inference |
| --- |
| |
| # Eliza-1 |
|
|
| Eliza-1 is the single elizaOS local-inference model repository. The v1 bundles are raw/base Qwen-lineage GGUF weights converted and packaged for the Eliza local harness; they are not fine-tuned. |
|
|
| Runtime bundles live under `bundles/<tier>/` so the app can resolve the catalog manifest and all component files from one repo. |
|
|
| APOLLO is the required optimizer for later fine-tuned releases. It keeps optimizer state small enough for full-parameter training on smaller GPUs, so publish scripts should not introduce a second optimizer path. |
|
|
| ## Bundle Matrix |
|
|
| | Tier | Remote path | Status | Files | Size | Voice note | |
| | --- | --- | --- | ---: | ---: | --- | |
| | 27b-256k | `bundles/27b-256k/` | ready | 55 | 20.94 GiB | policy satisfied or not declared | |
|
|
| Quantization policy: Q4_K_M is the default published runtime artifact today; Q6_K and Q8_0 are tracked in the app catalog as higher-precision variants for the hardware optimizer to select when those files are published. |
|
|
| ## Voice Modules |
|
|
| The voice sub-models live under `voice/<sub>/`. The unified layout |
| replaces the legacy `elizaos/eliza-1-voice-*` per-component repos |
| (consolidated 2026-05-15). Each sub-path carries a `provenance.txt` |
| linking back to its source repo at the time of consolidation. |
|
|
| | Sub-path | Contents | Source repo (deleted) | |
| | --- | --- | --- | |
| | `voice/omnivoice/` | OmniVoice TTS — `omnivoice-base-q4_k_m.gguf`, `omnivoice-base-q8_0.gguf`, `omnivoice-tokenizer-q4_k_m.gguf`, presets including `voice-preset-same.bin` and `voice-preset.elz2` (same-voice v0.1) plus README/eval/manifest | `elizaos/eliza-1-voice-omnivoice-same-v01` | |
| | `voice/emotion/` | Wav2Small emotion classifier — `wav2small-cls7-int8.onnx` (cls7 head) plus its config JSON, `eval.json`, `manifest.json`, README | `elizaos/eliza-1-voice-emotion` | |
| | `voice/turn/` | LiveKit turn detector — `intl/turn-detector-intl-q8.gguf` (281 MB), `intl/model_q8.onnx` (262 MB), tokenizer assets, README, manifest | `elizaos/eliza-1-voice-turn` | |
| | `voice/asr/` | Reserved (legacy repo was empty); ASR weights live under the canonical `voice/asr/` path: `eliza-1-asr-q8_0.gguf`, `eliza-1-asr-mmproj.gguf` | `elizaos/eliza-1-voice-asr` | |
| | `voice/kokoro/` | Kokoro TTS — extra preset `voices/af_sam.bin` (the canonical Kokoro weights live alongside under `voice/kokoro/kokoro-v1.0-q4.onnx` and `voices/af_bella.bin`, `voices/af_same.bin`) | `elizaos/eliza-1-voice-kokoro` | |
|
|
| Other voice components (`voice/diarizer/`, `voice/wakeword/`, `voice/speaker-encoder/`, `voice/vad/`, `voice/embedding/`, `voice/turn-detector/`, `voice/voice-emotion/`) were already shipping under the unified repo and remain unchanged. |
|
|
| Repository: `elizaos/eliza-1` |
|
|
