Voxyne
A tiny (128.8M-param) byte-level conversational language model with the
RahuKetu sigma_K control channel. It runs offline on CPU, and quantized to
int4 it stays coherent in ~80 MB — the intended default.
Scope — please read. Voxyne is built to converse, not to know. It has an identity and conversational ability; it is not a knowledge base and will make up facts if asked — knowledge is meant to come from external tools/retrieval (a separate system). Judge it on coherence, not factual recall.
Files — int4 is the default
| file | size | use |
|---|---|---|
voxyne-int4.onnx |
81 MB | DEFAULT — CPU/edge deployment (ONNX Runtime) |
voxyne-int8.onnx |
129 MB | int8 ONNX |
voxyne-fp32.onnx |
511 MB | full-precision ONNX |
voxyne-v0.1.pt |
258 MB | bf16 PyTorch weights (for the voxyne package / fine-tuning) |
The ONNX graphs are a single-token decode step with an explicit KV cache
(inputs h, pk, pv -> outputs logits, npk, npv); the byte embedding and
the sigma_K encoder run outside the graph. See the voxyne package for the
decode protocol.
Quick start (PyTorch)
pip install voxyne # then download voxyne-v0.1.pt from this repo
from voxyne import VoxyneConfig, build, load_weights, generate
model, enc = build(VoxyneConfig())
load_weights(model, "voxyne-v0.1.pt")
print(generate(model, enc, "who are you?", device="cpu"))
# -> "I'm Voxyne, an AI assistant created by Ramakrishnan."
Training-data provenance (why the license)
Voxyne's weights are trained on a mix that includes non-commercial sources, so the weights are released under CC BY-NC 4.0 (non-commercial). Free for research, education, and personal use.
| stage | sources (examples) | license note |
|---|---|---|
| pretrain | FineWeb-Edu, Cosmopedia, TinyStories | permissive / synthetic |
| grammar | WordNet, FrameNet, GoEmotions | permissive (attribution) |
| commonsense | ConceptNet, ATOMIC | ConceptNet = CC BY-SA |
| dialogue | SODA, UltraChat, OASST2, daily_dialog, empathetic_dialogues, WildChat | daily_dialog / empathetic = CC BY-NC; WildChat = AI2 ImpACT |
| identity | author-written | original |
The code (voxyne package) is Apache-2.0; only the weights are NC. A future
clean-data retrain (kalki) will carry Apache-licensed weights.
AI-assistance disclosure
Built by Ramakrishnan (ORCID 0009-0006-0905-7275). AI tools assisted with the
training/quantization automation and tooling; the model design, direction, and
all decisions are the author's.
License
Weights: CC BY-NC 4.0 (non-commercial). Code: Apache-2.0.