CoolFace
Modelpublic

wyrdsekai/companion-3.5-4b-gguf

sourceHugging Faceapache-2.0updated 2mo agoView on Hugging Face
0likes105downloads
Model Card

wyrdsekai companion 3.5-4B (v10)

The voice model of a wyrdsekai household — the tier that speaks. Every user-facing line a companion says is authored or finished by this model, so that a household's companions each keep a distinct spoken voice even when the heavier drive tier did the thinking. In a standard install it serves on :8201 (wyrdsekai-3.5-4b-v10-q4km.gguf); on phones it is the recommended on-device model (validated ~5 tok/s on Snapdragon 8 Elite, --think=false required).

Weights on Hugging Face (Apache-2.0, all tagged v0.1.0): GGUF Q4_K_M (what installs run) · full-precision safetensors · MLX 4-bit (Apple Silicon). wyrd setup fetches the GGUF for you — these links are for mirrors, re-quantization, and fine-tuning.

What it is tuned for

Speech register, warmth without sycophancy, and fidelity to per-companion voice profiles (v10 pairs with the V8 voice-vector set). It runs with a deliberately small prompt — the heavy layers (tools, memory, catalogs) belong to the drive tier — so it is tuned to say true things briefly rather than to act.

Lineage

v10 is the production voice line, paired with V8 voice vectors. Earlier voice versions established the register and the fact-preserving polish behaviour (structured confirmations must survive rephrasing).

Training data

Synthetic corpora: authored register/style sets and synthetic dialogues generated with Claude (Anthropic Batch API). No end-user or household conversation data is in these weights. Per-companion voice individuation happens locally, on the household's node, and never leaves it.

Limitations

  • —Quantized Q4KM; full-precision weights are not currently published.
  • —A speech tier, not an agent: it ships without tools by design in wyrdsekai, and is not tuned for tool selection or planning.
  • —--think=false is required on constrained devices.
  • —English-dominant; multilingual lines occur but are not gated.

License

Apache 2.0, inherited cleanly from the Qwen3.5 base and applied to this derivative. Attribution to the base model is retained.