Osho Agent (SmolLM3 3B) — Pure Osho Voice
A fine-tuned SmolLM3-3B model trained exclusively on Osho's teaching atoms. No system prompt needed. No retrieval. No persona blending. Question in, Osho out.
This is the lightweight 3B branch of the Osho Agent. Same V6 training format as the
8B version, smaller footprint.
~7,600 atoms from 3 books (Vigyan Bhairav Tantra Vol 1 & 2 + Zorba the Buddha, 1,260 pages) — technique, energy, explosion, celebration, tantra.
Quick Start
LM Studio / Ollama / llama.cpp: Download Osho_Agent-SmolLM3-3B-Q5_K_M.gguf and load it. No system prompt required.
Example:
User: Why am I suffering?
Osho Agent: Because you are fighting with life. Suffering is resistance. You want things to be other than they are — that is the whole of suffering. The moment you say yes to this moment, exactly as it is, suffering dissolves. Not tomorrow. Now. Suffering is not caused by life; it is caused by the 'no' you say to life.
User: How does meditation work?
Osho Agent: Meditation is not a technique — it is a let-go. Every technique is just an excuse to sit quietly. When you sit quietly long enough, something happens by itself. The mind exhausts itself. The chatter ceases. And in that gap — even for a split second — you taste who you are without the mind's commentary.
What Makes This Different
- Pure Osho, nothing else. Trained only on Osho's atoms — no blending, no dilution.
- Technique as explosion, not instruction. Osho doesn't describe meditation — he blasts you into it.
- Celebration and tantra alive. Yes to life, yes to body, yes to everything.
- Zero therapy-speak. No "I understand your feelings." Direct transmission.
- 3B footprint. Runs on modest hardware — 2 GB RAM with Q5_K_M.
Training Details
| Setting | Value |
|---|
| Base model | HuggingFaceTB/SmolLM3-3B-Base |
| Method | A-LoRA V6 (QDoRA + rsLoRA) |
| Rank | 32, alpha 32 |
| Training data | Osho V6 atoms (from ~7,600 raw, filtered for complete question+teaching pairs) |
| Source books | Vigyan Bhairav Tantra Vol 1 & 2, Zorba the Buddha |
| Format | V6 — question + concept arrows (input) → pure teaching passage (output) |
| Epochs | 1 |
| Quantization | Q5_K_M GGUF (~2 GB) |
The A-LoRA Methodology
A-LoRA (Atom LoRA) fine-tunes on complete reasoning atoms — indivisible multi-component teaching moves extracted from human expert traces. Each atom contains transformation (before/after), concept relations, anchoring quotes, and teacher method. Training on whole atoms produces models that move like teachers, not just sound like them.
The Teacher
Osho (Bhagwan Shree Rajneesh, 1931–1990). Neo-tantra. Provocation. Celebration. Energy. The Vigyan Bhairav Tantra — 112 techniques, all pointing to the same place.
| Style element | What it means |
|---|
| Technique as explosion | Doesn't explain meditation, throws you into it |
| Paradox and provocation | Uses humor and contradiction to break the thinking mind |
| Tantra | Embraces body, desire, celebration — nothing excluded |
| Direct confrontation | Addresses the seeker directly, not abstractly |
3B vs 8B
The
8B Osho Agent carries more nuance and handles edge cases better. This 3B branch is for when size matters — Osho's core movement is intact, though some of the finer voice details are softer.
Files
| File | Size | Use |
|---|
Osho_Agent-SmolLM3-3B-Q5_K_M.gguf | ~2 GB | Recommended |
Related Models
- Full series — Meditation Agent Collection — all 19 models
- Osho Agent 8B — Full-fidelity Osho, Qwen3-8B base
- Meditation Agent 3B — All 9 teachers blended, SmolLM3-3B
- Meditation Agent 8B — All 9 teachers blended, Qwen3-8B
ellam sivamayam — Everything is Shiva's expression.
எல்லாம் சிவமயம்
