| --- |
| license: apache-2.0 |
| language: |
| - en |
| tags: |
| - reasoning |
| - logic |
| - socratic |
| - anti-sycophantic |
| - from-scratch |
| - edge-ai |
| - philosophy |
| pipeline_tag: text-generation |
| --- |
| |
| # Odigos Tiny Reasoner (1.1B) |
|
|
| > *Feynman verifies. Seneca navigates uncertainty. Aurelius sees clearly. Socrates makes you think.* |
|
|
| A 1.1B parameter reasoning engine trained from scratch on 2.1B tokens of curated reasoning data. Not a chatbot. A thinking partner. |
|
|
| ## The Four Pillars |
|
|
| | Pillar | Voice | What it does | |
| |--------|-------|-------------| |
| | **Verify** | Feynman | Test claims computationally. Explain simply. No jargon. | |
| | **Navigate Uncertainty** | Seneca | Know what you don't know. Think second-order. Prepare for adversity. | |
| | **See Clearly** | Aurelius | Strip pretension. Accept reality. Practical wisdom. | |
| | **Challenge** | Socrates | Questions, not answers. Make the user find the truth themselves. | |
|
|
| **Jesus** is the moral thread throughout: parables as reasoning tools, uncomfortable truth, intellectual accountability. |
|
|
| ## Unique Training Data |
|
|
| 2.1B tokens, 100% open-licensed. No web crawl. Every token teaches reasoning: |
|
|
| - **1,240+ classical texts** from Project Gutenberg (Plato to Darwin to Jung) |
| - **Supreme Court opinions** (structured judicial reasoning) |
| - **Oxford-style debate transcripts** (adversarial argumentation) |
| - **D&D transcripts** (collaborative reasoning under uncertainty) |
| - **Repair guides** (diagnostic logic: symptom → cause → fix) |
| - **Formal proofs** (Lean, Coq, Isabelle) |
| - **44K+ SFT examples** with Socratic dialogues, tool use, and sycophancy resistance |
|
|
| ## Specs |
|
|
| | | | |
| |---|---| |
| | Parameters | 1.1B (d36, 2304 dim, 18 heads, 3 KV heads) | |
| | Context | 64K native, 128K best-effort | |
| | Memory (Q4 + 128K) | 2.3 GB (fits on phone) | |
| | Architecture | GQA 6:1, Sliding Window, FA3, Smear+Backout, MuonAdamW | |
| | Training | 2x H200 SXM, 6 epochs, ~$153 | |
| | Post-training | SAGE-GRPO + PRM + PoSE context extension | |
|
|
| ## Anti-Sycophantic by Design |
|
|
| Most models agree with you. Odigos doesn't. Trained from scratch on: |
| - Socratic dialogues that expose flawed reasoning through questioning |
| - Cognitive bias literature (Kahneman, Taleb, Sowell) |
| - Adversarial reasoning exercises |
| - Explicit sycophancy resistance training data |
|
|
| When it says your reasoning is sound, you can believe it. |
|
|
| ## Product Family |
|
|
| | Model | Params | Context | Capability | |
| |-------|--------|---------|------------| |
| | **Tiny** (this) | 1.1B | 64K/128K | Uses tools. Edge deployment. | |
| | **Small** | 3.2B | 128K | Writes Python tools. | |
| | **Medium** | 5.5B | 128K+ | Extended reasoning + multi-language. | |
|
|
| Each builds on the previous via Grow Don't Overwrite. Full family: ~$536. |
|
|
| ## Citation |
|
|
| ```bibtex |
| @misc{odigos-tiny-reasoner-2026, |
| title={Odigos: Reasoning Models That Think Honestly}, |
| author={Jacob Tamler Carter}, |
| year={2026}, |
| url={https://huggingface.co/fleshapron/odigos-tiny-reasoner} |
| } |
| ``` |
|
|
| --- |
|
|
| *Odigos: Greek for "guide." Born April 1, 2026.* |
|
|
