i14 Labs Launches Sugar Glider LLM Architectures With O(1) Inference

i14 Labs launched Sugar Glider, a new family of LLM architectures delivering frontier performance with extreme memory and compute efficiency, tested up to roughly 1 trillion parameters. The architecture provides O(1) per-token inference and a fixed-size KV cache, achieving 100% single-needle retrieval at 1 billion tokens and up to 32x more requests per server at one million tokens. Early access is available via a waitlist.
Event version 1.
Key insights
Efficiency at frontier scale
Sugar Glider delivers a new Pareto frontier of extreme memory and compute efficiency while maintaining frontier performance, tested up to roughly 1 trillion parameters.
Constant-cost inference
The architecture provides O(1) per-token inference and a fixed-size KV cache as context grows.
Billion-token retrieval results
Sugar Glider achieved 100% single-needle retrieval and 90.5% multikey retrieval at 1 billion tokens of context.
Serving gains at one million tokens
At one million tokens, i14 reports up to 32x more requests per server than its baseline model, and memory size can be chosen at deployment without retraining.
Open-weight model conversion
Several powerful and popular open-weight models have been converted to O(1) inference and fixed memory, with each conversion taking just hours.