# Saturn Cloud Brings Tileward’s Compressed Models to Its Token Factory, Turning the Same GPUs Into More Tokens Page: https://tileward.com/blogs/saturn-cloud-partnership/ Press release. Tileward’s compressed open models free up GPU memory, so operators on Saturn Cloud’s token factory platform serve more tokens from the GPUs they already own, with no change to their serving stack or rate cards. NEW YORK, Nov. 20, 2026 -- [Saturn Cloud](https://saturncloud.io/), the inference platform for enterprises running open models, today announced a partnership with [Tileward](https://tileward.com/) to bring its compressed open models to Saturn Cloud’s token factory platform. Operators run the same models in less GPU memory, so each GPU serves more tokens. Saturn Cloud’s token factory turns GPU fleets into managed inference and fine-tuning services that operators sell to their own customers, under their own brand. Tenant isolation, quotas, metering, and enterprise governance are built in. It runs on [NVIDIA Dynamo](https://saturncloud.io/blog/running-a-token-factory-on-nvidia-dynamo/), in public cloud, private cloud, or on-premises. Serving Kimi K3 to 16 live agent sessions on NVIDIA GB300 GPUs, Saturn Cloud delivered [7,380 total tokens per second per GPU](https://saturncloud.io/#performance) at about 37 tokens per second per user, matching the SemiAnalysis InferenceX benchmark recipe on the same GPUs. Tileward’s models run on this serving stack. “Operators sell inference from GPUs they’ve already paid for, and memory is usually what runs out first,” said Sebastian Metti, Founder and CEO of Saturn Cloud. “Tileward gives that memory back, so operators can take on more customers and longer contexts with the fleet they have, without buying more GPUs.” Tileward, a member of the [NVIDIA Inception](https://www.nvidia.com/en-us/startups/) program for startups, offers compressed models that are up to 78% smaller than the published weights. Tileward evaluates each one against the original on benchmarks including MMLU, SWE-bench Verified, and AgentBench. “Making a model smaller was never the hard part. Proving it still gives the right answers is,” said Anantha Sharma, Founder and CTO of Tileward. “We publish how we measure every model, so operators on Saturn Cloud can check the quality before they sell a single token of it.” Operators swap in Tileward’s weights for the published ones, and their serving stack, tenants, rate cards, and brand stay as they are. The freed memory lets them run more copies of each model, fit large models on fewer GPUs, and serve more concurrent users with longer contexts. Serving more tokens from each GPU lowers the cost per token, giving operators more margin on input, cached, and output rates, or room to price lower. ## Available Now on Saturn Cloud Compressed versions of gpt-oss-120b, GLM-5.3, and DeepSeek-V4-Flash are available now to neoclouds, AI factory operators, telcos, sovereign AI providers, and enterprises running Saturn Cloud’s token factory platform on their own GPUs. Contact [Saturn Cloud](https://saturncloud.io/) or [Tileward](https://tileward.com/) to add them to a deployment. ## About Saturn Cloud Saturn Cloud is the inference platform for enterprises running open models. Teams run open models and agents on dedicated GPUs from neocloud partners or inside their own networks, with performance backed by an SLA. GPU operators use Saturn Cloud’s token factory platform to sell inference from their fleets. Learn more at [saturncloud.io](https://saturncloud.io/). ## About Tileward Tileward makes open models smaller while preserving publisher accuracy and runs them as a hosted API, in a company’s own cloud, or fully air-gapped. Learn more at [tileward.com](https://tileward.com/). ## Media contact Saturn Cloud: [mel@saturncloud.io](mailto:mel@saturncloud.io) Tileward: [marketing@tileward.com](mailto:marketing@tileward.com)