Model Releases
GrootTeam Releases Groot Family of Open-Weight MoE Reasoning Models
GrootTeam introduces a new family of four Mixture-of-Experts reasoning models, ranging from 19B to 280B parameters, optimized for high-efficiency complex reasoning.
October 23, 2026
Model ReleaseGroot
Introduction
On October 23, 2026, a new player in the open-weights ecosystem, GrootTeam, disrupted the reasoning model landscape with the release of Groot. This isn't just another incremental update; it is a specialized family of four Mixture-of-Experts (MoE) models designed specifically to tackle complex, real-world reasoning tasks with unprecedented efficiency.
What makes Groot particularly striking is its departure from the 'scale-at-all-costs' philosophy. Instead of training massive models from scratch, Groot utilizes a highly efficient post-training lineage, leveraging Qwen3.6-35B-A3B as a base. By using fewer than 0.1B tokens of high-quality open-source data, Groot demonstrates that capability per parameter can significantly outperform raw scale.
- Release Date: October 23, 2026
- License: Apache-2.0
- Architecture: Mixture-of-Experts (MoE)
- Primary Focus: Complex reasoning and agentic workflows
Key Features & Architecture
The Groot family is comprised of four distinct model sizes, allowing developers to choose the right balance of performance and computational cost for their specific deployment needs. The architecture utilizes a Mixture-of-Experts approach, ensuring that only a fraction of the total parameters are active during any single inference pass, which drastically reduces latency.
The models are designed for massive context windows and high-density reasoning. From the lightweight 19B model to the powerhouse 280B variant, the family supports extensive context lengths, making them ideal for long-document analysis and complex codebase navigation. Furthermore, the models exhibit remarkable token efficiency, producing much shorter, more concise outputs than their competitors without sacrificing reasoning depth.
- 19B-A3B: Optimized for edge and consumer hardware
- 65B-A6B: The balanced mid-range performer
- 238B-A18B: High-capacity agentic reasoning with 1M-token context
- 280B-A16B: The flagship model for complex software engineering tasks
- Efficiency: Significant reduction in output token counts compared to Qwen and Gemma
Performance & Benchmarks
The empirical data released by GrootTeam suggests a paradigm shift in efficiency. The smallest model, Groot-19B-A3B, remarkably outperforms both Qwen3.6-35B-A3B and Gemma-4-26B-A4B on 12 of 16 key benchmarks. Despite having 45.7% fewer total parameters than the Qwen model, it achieved an unweighted average score of 41.44, compared to Qwen's 38.08 and Gemma's 30.05.