Sugar Glider

Sugar Glider is a constant-memory LLM family from Melbourne-based i14 / i14labs, built on the Cerbearus architecture with O(1) memory and compute per token as context grows.

Sugar Glider 120B is positioned for about 1M context while keeping a fixed-size memory footprint—on the order of ~1/32 of a parent KV cache at 1M—so the same GPUs can serve far more concurrent long-context requests. The family also includes a 32B SKU, is built on Nemotron bases, and offers a runtime-configurable memory budget. Product pages live at https://www.i14.ai/models with the 120B news post at https://www.i14.ai/news/sugar-glider-120b .

Highlights:

• Constant-memory long-context architecture (Cerbearus); O(1) memory per token;
• ~1M context on the 120B SKU with fixed-size memory (~1/32 parent KV at 1M as stated);
• Runtime-configurable memory budget; 120B and 32B SKUs;
• Built on Nemotron bases; models hub https://www.i14.ai/models ;
• Official launch materials from i14 (Melbourne).

Why it matters

Standard transformers pay linear (or worse in practice) KV-cache cost as context length grows, which caps concurrent long-document sessions on a fixed GPU pool. Sugar Glider’s constant-memory / O(1)-per-token claim attacks that economics: keep ~1M context while holding a fixed memory budget, then pack more simultaneous long-context jobs on the same hardware.

For agent and RAG products that routinely pass large corpora or long tool traces, a runtime-configurable memory knob plus 120B/32B SKUs is more actionable than “infinite context” marketing. Radar kept the term because i14 shipped an official news post, whitepaper-style materials, and a models page in this window.

Limits

Constant-memory designs trade exact full-attention fidelity for bounded state—validate quality on your long-context evals, not only tokens-per-dollar. The ~1/32 KV comparison is i14’s stated figure at 1M; reproduce on your serving stack. Confirm license, Nemotron-base obligations, region availability, and current API/model ids on https://www.i14.ai/models . Do not assume unpublished parameter breakdowns or private latency SLOs.

1. Read https://www.i14.ai/news/sugar-glider-120b and browse https://www.i14.ai/models for SKUs and docs;
2. Pick 120B vs 32B based on quality needs, then set the runtime memory budget for your concurrency target;
3. Benchmark long-context RAG/agent traces against a full-KV baseline on the same GPUs;
4. Verify license and Nemotron-base terms before production; re-check context and memory claims on the current release notes.