Flagship model
Snort-1
Our most capable model to date. Snort-1 is trained to be helpful, harmless, and horizontal.
Overview
- Release
- Private preview, Q3 2026
- Context window
- 128k sniffs
- Parameters
- 11 billion wrinkles
- Modalities
- Sniff, bark, head tilt
- Training data
- 40 petabelly-rubs
- Pricing
- $0.40 per million input treats · $1.60 per million output treats
Benchmarks
Snort-1 sets a new state of the art on four of five evaluations. We continue to investigate the fifth.
Higher is better. Scores in percent.
| Benchmark | Snort-1 | Leading Cat Model | Previous SOTA (a tennis ball) |
|---|---|---|---|
| MMPU · Massive Multitask Pug Understanding | 91.4% | 74.2% | 8.6% |
| ARC-Bark · Abstract Reasoning, Canine | 88% | 61.5% | 12% |
| HumanSnack · Snack Acquisition from Humans | 96.2% | 83.9% | 3.1% |
| SNIFF-bench · Olfactory Retrieval at Distance | 99.1% | 97.8% | 0.4% |
| TableKnock-Hard · Gravity-Based Object Relocation | 12% | 99% | 0% |
Scaling
Performance improves predictably with training treats. New capabilities emerge at specific thresholds, the first of which we have observed at roughly 40 treats.
Model card
- Intended use
- Being held. Secondary uses include lap occupancy and floor-based reasoning.
- Out-of-scope use
- Agentic mailman interception. Stair navigation.
- Training
- Pretrained on 40 petabelly-rubs, then tuned with reinforcement learning from treat feedback (RLTF).
- Known limitations
- Stairs. Performance degrades in the vicinity of open windows and vacuum cleaners.
- Evaluation
- All results were produced by Snort-1 under supervised conditions (a blanket).