Canopy-258M-R3 v2 (Browser Agent & Unified Multi-Domain Flagship)
Canopy-258M-R3 v2 is an ultra-efficient 258.56M parameter Recurrent Mixture-of-Experts (MoE) causal language model designed for high-speed edge computing, mathematical reasoning, Python code synthesis, and browser use.
What is New in v2
- Unified Multi-Domain Mastery: 50.0% HumanEval executed pass rate, 100% code syntax, balanced factual and tool execution.
- Natural Browser Interaction Kinematics: Cubic Bézier mouse paths, Gaussian typing cadence, and Set-of-Marks coordinate grounding.
- Visual Screenshot Auditing: Step-by-step visual audit trail generation with coordinate crosshairs.
- Extended Browser Suite Benchmark: 100% pass rate across 7 distinct real-world web interaction modalities.
Model Architecture
- Parameters: 258,555,654 total (~112M active per token)
- Routing: Top-2 across 8 specialized experts with Tokenwise Thought Bus
- Recurrence: 6 middle MoE blocks executed twice (18 effective layers)