Tripplet Research

An independent research collective exploring the frontier of open language models through weight-space merging, efficient fine-tuning, and architectural experimentation.

We focus on practical questions: how far can small-to-mid-scale models (7B–13B) be pushed on consumer hardware without sacrificing reasoning, tool use, or multilingual ability?

Released Models

🧠 Suzhou 3.1

An 8.95B SLERP merge combining agent/tool-use capabilities with a broad general-purpose base. Built on the Qwen3.5 architecture with mixed linear + full attention layers, 262K context window, and support for 29+ languages.

• Method: Gradient-blended SLERP across attention and MLP layers
• Base models: kai-os/Carnice-9b + principled-intelligence/Qwen3.5-9B-text-only
• License: Apache 2.0

Research Directions

• Weight-space merging — SLERP, TIES, DARE, and layer-wise gradient blending across differently-tuned checkpoints
• Efficient reasoning — Chain-of-thought quality in 9B-class models with long context
• Edge-capable agents — Tool-using models that fit on consumer GPUs and Apple Silicon
• Multilingual generalization — Understanding how merging affects non-English performance

Philosophy

We believe the most interesting capability gains at this scale come not from training larger models, but from recombining the weights of models that already know different things. Merging is cheap, repeatable, and surprisingly effective — and it deserves more rigorous study.

License

All Tripplet Research releases are Apache 2.0 unless explicitly noted otherwise.

Tripplet Research ships under the Tripplet Models brand. Interested in collaborating? Open a discussion on any of our repos.