OusiaResearch / Aureth_Qwen3.5‑0.8B — a lean, efficient fine-tune of Qwen3.5‑0.8B, purpose-built for light inference workloads and resource‑constrained deployments.
∴ Model Summary
Field
Detail
Base model
Qwen/Qwen3.5‑0.8B
Parameters
0.9 B (full precision)
Context length
4 096 tokens
Training volume
65 M tokens
Precision
BF16 safetensors + FP32
License
Apache 2.0
Framework
Unsloth + HuggingFace TRL
Aureth — from the Ouroboros axiom ουσία (ousia, "being" or "essence") — denotes a model stripped to its functional core: efficient, present, and ready to run where larger models cannot.
◦ Training
Corpus
Single‑phase full fine‑tuning on a curated mix of:
Instructional dialogue — structured for clarity and concision
Domain‑neutral conversational data — broad coverage over specialist depth
The 65 M token volume is intentionally bounded — calibrated for lightweight adaptation without memorisation or alignment degradation.
Objective
Standard next‑token prediction (cross‑entropy) at full precision (bf16 → fp32) — no adapter layers, no quantisation. The full parameter space is updated at modest learning rate to preserve the base model's innate capabilities while gently steering tone and instruction‑following.
Hardware
Trained via Unsloth Studio — 2× faster than naive bf16 fine‑tuning with equivalent or better loss convergence.
Aureth‑0.8B is not a reasoning model. It excels at:
Simple Q&A and information retrieval
Short‑context summarisation
Light conversational agents
Scripted automation / tool‑calling at low branching depth
Running as a always‑on personal assistant on modest hardware (MacBook, single‑GPU workstation, CPU inference via llama.cpp/MLX)
◦ Ousia Research
Aureth is the production fine‑tune lineage of the Ousia Project — a family of models trained on synthetically generated cognitive traces with the goal of developing coherent, consistent, and ethically grounded language models.
Variant
Context
Corpus
Status
Aureth‑Qwen3.5‑0.8B
4 K
65 M tokens
◉ Public
Aureth‑Qwen3.5‑4B
4 K
~262 M tokens
∙ Training
Aureth‑Qwen3.5‑9B
4 K
TBD
∙ Planned
"The smallest Ouroboros still eats its own tail."
Model card v1.0 — Ousia Research — April 2026Licensed under Apache 2.0 — OusiaResearch / Aureth_Qwen3.5‑0.8B
