# i14 Sugar Glider 120B: Million-Token Context With Constant Memory

i14's Sugar Glider 120B targets long-context serving with fixed memory. Its capacity claims, RULER trade-offs, API access and illustrative pricing explained.

By Paras Madan | 2026-10-21


i14 announced Sugar Glider 120B on October 20, a language model aimed at making long-context inference less demanding on server memory. The company says it can process a million tokens while keeping the memory carried between tokens at a fixed size.

That is useful for teams building document-analysis tools or agents with long histories. The important distinction is between fitting more requests in memory, answering them quickly and answering them accurately. The release provides evidence about those questions at different levels of detail.

## What constant memory means here

The [launch announcement](https://www.i14.ai/news/sugar-glider-120b) places Sugar Glider within i14's Cerbearus architecture family. Rather than letting the stored context grow with each token, the model maintains a fixed-size memory. Developers can select its size when running the model without retraining the weights.

A fixed memory budget makes capacity planning simpler in principle: a long request does not keep demanding a larger context store as it grows. It does not make an entire million-token input free to process. The model still has to read the tokens, and the capacity of its memory is a different question from the length of the input it accepts.

For a product team, this is a serving trade-off worth measuring. A smaller state may let a machine keep more sessions active, but a useful deployment still needs acceptable response times and reliable recall of the information each session contains.

## What the 32-times capacity claim covers

The [Cerbearus whitepaper](https://www.i14.ai/papers/cerbearus-whitepaper.pdf) identifies the 120B model's parent as NVIDIA Nemotron 3 Super 120B-A12B. On four H200 GPUs, it reports capacity for 453 simultaneous million-token requests, compared with fourteen for the parent. On four H100s, the corresponding counts are 73 and two.

Those figures describe requests that fit in memory. They do not establish a 32-times increase in tokens per second, a matching latency improvement or a 32-times reduction in a customer's bill.

The paper also evaluates a larger memory setting. At four times the default memory budget, the H200 configuration fits 141 million-token requests, about ten times the parent. This gives operators a capacity setting to investigate rather than a single fixed configuration.

## The long-context quality trade-off

i14 reports a 73.3 average on the full thirteen-task RULER benchmark at one million tokens. The whitepaper's comparison chart places the parent at 95.0 at that length. At 32,000 tokens, the scores are much closer: 95.7 for Cerbearus-120B and 96.3 for the parent.

That distinction should stay attached to the capacity claim. The paper's broader statements about preserving capabilities should not be read as equal long-context retrieval performance at a million tokens.

[RULER](https://arxiv.org/abs/2404.06654) tests more than finding a single fact in a long passage. It includes multiple-needle retrieval, tracing chains of information and aggregation. Even so, it is a synthetic evaluation. A team handling repositories, research files or support histories should add tests that resemble those actual inputs and the answers users need.

## API access and pricing need separate checks

The [model documentation](https://www.i14.ai/docs/models/sugar-glider/) describes a text-only thinking model served through an OpenAI-compatible API. Its documented identifier is i14-sugarglider-nemotron3-super-120B-v1. Image input is not supported.

Thinking tokens count toward the output-token budget. If that budget is exhausted during reasoning, the response can end without a final answer. The docs also recommend streaming for longer requests because a non-streaming response taking more than about thirty seconds can encounter a gateway timeout.

The [model directory](https://www.i14.ai/models) lists 120 billion total parameters, twelve billion active parameters and a one-million-token context. It displays input and output figures of $0.10 and $0.40 per million tokens, but explicitly labels Sugar Glider prices as illustrative. They should not be treated as a firm commercial quote.

The announcement directs users to an early-access waitlist or a conversation about private deployment. Cloud, VPC and on-premises options are described; that is different from a public weights download or guaranteed immediate access.

For an initial evaluation, ask for the intended deployment configuration and confirmed pricing, then measure completed tasks at several context lengths. The useful result is the balance of recall, latency and cost under your workload, rather than the largest advertised context alone.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders explain model capabilities and infrastructure trade-offs through clear launch stories and developer content.

[Book a 30-minute call](https://cal.com/paras-madan/30min).