OusiaResearch/Aureth_Qwen3.5-0.8B

Aureth Qwen3.5‑0.8B

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:

• Synthetic cognitive traces — Hermes‑4.3‑36B‑generated reasoning chains
• 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.

---

○ Intended Use

 lightest workloads · embedded hosts · edge devices
 rapid prototyping · 24‑7 low‑latency inference

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 2026 Licensed under Apache 2.0 — OusiaResearch / Aureth_Qwen3.5‑0.8B