Open models at full speed, on dedicated GPUs
Choose from a large catalog of open models and agents, or bring your own. We run them on dedicated GPUs from our neocloud partners, tune each deployment for your traffic and put the performance in your contract.
Self-service has limited GPU availability. Contact us for more capacity.
Your agents and applications
Open models
Dedicated GPUs from our neocloud partners
Speed
We profile your workload and tune every deployment for it, so each endpoint gets more tokens out of its GPUs.
Capacity
We partner with neoclouds bringing new GPU capacity online, before it reaches the open market. That lets us offer GPUs for less than you will find elsewhere. You get an endpoint, with no GPUs to find or run.
Model choice
Run Kimi K3, GLM-5.3, DeepSeek V4 Flash, gpt-oss-120b and many more from our catalog, or bring your own model. The same goes for agents.
Fine-tuning
Train on your own data, on models that support it, and serve the result from your dedicated endpoint.
Control
SOC 2, GPUs dedicated to your endpoint and, when you need it, deployment inside your own network, including air-gapped.
Benchmark performance, with enterprise orchestration on top
You don't give up inference performance to get a managed, secure platform. On Kimi K3, Saturn Cloud matched the InferenceX benchmark on the same GPUs. On GLM-5.3, profiling more than doubled throughput at 16 live sessions.
Kimi K3: level with the InferenceX benchmark
Total tokens per second per GPU, 16 live agent sessions, same GB300 GPUs
Built on the NVIDIA Dynamo serving stack, with no proprietary engine to lock you in. We reached the benchmark on the same GPUs by tuning for the model, and we do the same for your workload.
GLM-5.3: 2.4x the throughput after tuning
Total tokens per second per GPU, by live agent sessions, same four GB300 GPUs
How we measured this
- Workload for both: recorded coding-agent traffic from InferenceX's AgentX benchmark, with input and output tokens counted and prefix caching on. About 99 percent of Kimi K3's counted tokens, and about 90 percent of tuned GLM-5.3's, are input, most of it reused context.
- Kimi K3: tensor parallel 8 over two GB300 trays. The reference is the InferenceX recipe run directly on our trays, with three multicast settings off because the partition's limit required it. Saturn Cloud ran vLLM 0.31.0 on Dynamo 1.6.0 with decode context parallel 8. Both forced speculative-decoding acceptance (3.62 on Saturn Cloud), and both ran at about 37 tokens/s per user. A default Kimi K3 endpoint, with no speculative decoding, gave 5,557 at 22.4 tokens/s per user on this test, and about 2,560 at 50 tokens/s per user. With acceptance forced, interpolating to 50 tokens/s per user gives about 6,060 on Saturn Cloud, against 6,083 published for the recipe on B300.
- GLM-5.3 trace: the public cc-traces-weka-062126, replayed with agentx-harness 1.0.12, each point 20 minutes. Our 4-GPU context limit is 524,288 tokens, so 70 of the 393 sessions were dropped.
- GLM-5.3 setups: starting setup is our catalog defaults at the time, on one endpoint. Tuned adds decode context parallel 4, a 256 GB CPU KV tier and MTP with 3 draft tokens, with totals averaged over two endpoints that agreed within 1 to 4 percent. At 16 sessions, per-user speed rose from 8.2 to 30.7 and 38.4 tokens/s.
- Your numbers will differ with your prompts and traffic.
We back it with a contract
Once your workload is profiled and benchmarked, we commit to the throughput and latency agreed for your workload, not just availability. We set targets for each deployment, based on your model, the GPUs it runs on, your request characteristics, concurrency and capacity. SLA commitments are backed by contractual remedies, such as service credits, where applicable.
- 1Profile your workload
- 2Optimize and benchmark
- 3Contractually guarantee performance
Example SLA
| Metric | Example guarantee |
|---|---|
| Throughput | ≥ 5,000 output tokens/sec |
| Time to first token (p95) | ≤ 2.0 seconds |
| Inter-token latency (p95) | ≤ 50 ms |
| Availability | 99.9% |
Illustrative numbers for a hypothetical deployment, not actual performance claims. Your targets are tuned to your model, its GPUs and your workload.
Our catalog, or your own model
Start from our large catalog of open models, or bring your own. These are some of the model families we run. Tell us what you need and how you will use it, and we will size the deployment for you.
Kimi
GLM
DeepSeek
MiniMax
Nemotron
gpt-oss
Qwen
Llama
Gemma
Mistral
Phi
Granite
Falcon
OLMo
Yi
SmolLM
See more of our catalog
- 01-ai/Yi-1.5-34B
- 01-ai/Yi-1.5-34B-Chat
- 01-ai/Yi-1.5-9B
- 01-ai/Yi-1.5-9B-Chat
- 01-ai/Yi-1.5-6B
- 01-ai/Yi-1.5-6B-Chat
- Qwen/Qwen3-235B-A22B
- Qwen/Qwen3-235B-A22B-Instruct-2507
- Qwen/Qwen2.5-72B
- Qwen/Qwen2.5-72B-Instruct
- Qwen/Qwen2.5-32B
- Qwen/Qwen2.5-32B-Instruct
- Qwen/Qwen3-32B
- Qwen/Qwen3-30B-A3B
- Qwen/Qwen3-30B-A3B-Instruct-2507
- Qwen/Qwen3-Coder-30B-A3B-Instruct
- Qwen/Qwen3.8-27B
- Qwen/Qwen2.5-14B
- Qwen/Qwen2.5-14B-Instruct
- Qwen/Qwen3-14B
- Qwen/Qwen3-14B-Base
- Qwen/Qwen3-8B
- Qwen/Qwen3-8B-Base
- Qwen/Qwen2.5-7B
- Qwen/Qwen2.5-7B-Instruct
- Qwen/Qwen3-4B
- Qwen/Qwen3-4B-Base
- Qwen/Qwen2.5-3B
- Qwen/Qwen2.5-3B-Instruct
- Qwen/Qwen3-1.7B
- Qwen/Qwen3-1.7B-Base
- Qwen/Qwen2.5-1.5B
- Qwen/Qwen2.5-1.5B-Instruct
- Qwen/Qwen3-0.6B
- Qwen/Qwen3-0.6B-Base
- Qwen/Qwen2.5-0.5B
- Qwen/Qwen2.5-0.5B-Instruct
- allenai/Olmo-3.1-32B-Instruct
- allenai/Olmo-3.1-32B-Think
- allenai/Olmo-3-7B-Instruct
- allenai/Olmo-3-7B-Think
- deepseek-ai/DeepSeek-R1
- deepseek-ai/DeepSeek-V3
- deepseek-ai/DeepSeek-V4-Flash
- deepseek-ai/DeepSeek-V4-Flash-0731
- deepseek-ai/DeepSeek-R1-Distill-Llama-70B
- deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
- deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
- deepseek-ai/DeepSeek-R1-Distill-Llama-8B
- deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
- deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
- google/gemma-4-31B
- google/gemma-4-31B-it
- google/gemma-2-27b
- google/gemma-2-27b-it
- google/gemma-3-27b-it
- google/gemma-3-27b-pt
- google/gemma-4-26B-A4B
- google/gemma-4-26B-A4B-it
- google/gemma-3-12b-it
- google/gemma-3-12b-pt
- google/gemma-4-12B
- google/gemma-4-12B-it
- google/gemma-2-9b
- google/gemma-2-9b-it
- google/gemma-3-4b-it
- google/gemma-3-4b-pt
- google/gemma-4-E4B
- google/gemma-4-E4B-it
- google/gemma-2-2b
- google/gemma-2-2b-it
- google/gemma-4-E2B
- google/gemma-4-E2B-it
- google/gemma-3-1b-it
- google/gemma-3-1b-pt
- HuggingFaceTB/SmolLM2-1.7B
- HuggingFaceTB/SmolLM2-1.7B-Instruct
- HuggingFaceTB/SmolLM2-135M-Instruct
- ibm-granite/granite-4.0-h-small
- ibm-granite/granite-4.0-h-tiny
- ibm-granite/granite-4.0-micro
- meta-llama/Llama-4-Maverick-17B-128E-Instruct
- meta-llama/Llama-4-Scout-17B-16E
- meta-llama/Llama-4-Scout-17B-16E-Instruct
- meta-llama/Llama-3.1-70B
- meta-llama/Llama-3.1-70B-Instruct
- meta-llama/Llama-3.3-70B-Instruct
- meta-llama/Llama-3.1-8B
- meta-llama/Llama-3.1-8B-Instruct
- meta-llama/Llama-3.2-3B
- meta-llama/Llama-3.2-3B-Instruct
- meta-llama/Llama-3.2-1B
- meta-llama/Llama-3.2-1B-Instruct
- microsoft/phi-4
- microsoft/Phi-3.5-mini-instruct
- MiniMaxAI/MiniMax-M3
- mistralai/Mixtral-8x22B-Instruct-v0.1
- mistralai/Mixtral-8x22B-v0.1
- mistralai/Mixtral-8x7B-Instruct-v0.1
- mistralai/Mixtral-8x7B-v0.1
- mistralai/Mistral-Small-24B-Base-2501
- mistralai/Mistral-Small-24B-Instruct-2501
- mistralai/Mistral-Nemo-Base-2407
- mistralai/Mistral-Nemo-Instruct-2407
- mistralai/Mistral-7B-Instruct-v0.3
- mistralai/Mistral-7B-v0.3
- moonshotai/Kimi-K3
- nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
- nvidia/Llama-3.1-Nemotron-70B-Instruct-HF
- nvidia/Llama-3.3-Nemotron-Super-49B-v1
- nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16
- nvidia/Llama-3.1-Nemotron-Nano-8B-v1
- openai/gpt-oss-120b
- openai/gpt-oss-20b
- tiiuae/Falcon-H1-34B-Instruct
- tiiuae/Falcon3-10B-Base
- tiiuae/Falcon3-10B-Instruct
- tiiuae/Falcon-H1-7B-Instruct
- tiiuae/Falcon3-7B-Base
- tiiuae/Falcon3-7B-Instruct
- zai-org/GLM-5.3
- zai-org/GLM-4.5
- zai-org/GLM-4.6
- zai-org/GLM-4.5-Air
- zai-org/GLM-4.5-Air-Base
Any open-source agent, set up for you
Bring your own agents or any open-source agent, for coding, research, documents, data and browser work. These are a sample of what we run. We curate a catalog for each customer and set up the agents you pick as part of your contract.
ClineAiderContinueOpenCodeTabbyAnythingLLMGPT ResearcherSTORM
Kotaemon
PandasAI
Nanobrowser
OpenClaw
RAGFlow
LibreChat
LightRAG
OpenManus
gptme
Open NotebookSee more agents
- Agent S
- AgenticSeek
- AI Data Science Team
- Aider
- AiderDesk
- AnythingLLM
- AutoDev
- BambooAI
- Cline
- CodeCompanion.nvim
- Continue
- Datus
- DB-GPT
- Deep Research
- DeepAnalyze
- DeepSearcher
- DeepTutor
- DocETL
- DocsGPT
- FinRobot
- Gito
- GPT Academic
- GPT Researcher
- gptme
- IronClaw
- Khoj
- Kotaemon
- LibreChat
- LightRAG
- Local Deep Research
- MAESTRO
- mini-SWE-agent
- MiniSearch
- Moltis
- nanobot
- Nanobrowser
- Nanocoder
- nao
- Octofriend
- Onyx
- Open Interface
- Open Notebook
- OpenClaw
- OpenCode
- OpenDeepWiki
- OpenManus
- Page Assist
- PandasAI
- Paperless-AI
- paperless-gpt
- PicoClaw
- ProxyAI
- QueryWeaver
- RAG Web UI
- RAGFlow
- Sparrow
- STORM
- SurfSense
- SWIRL
- Tabby
- twinny
- Vane
- Yuxi
- ZeroClaw
Logos belong to their owners and identify each project. Listing one does not imply endorsement.
A deployment shaped around your traffic
A dedicated endpoint has its own GPUs, its own worker layout and its own scaling limits, so other customers' traffic does not compete with yours.
- Run prefill and decode together or in separate pools, each with its own GPU size.
- Set the minimum workers and the limit it can scale to.
- Favor latency, throughput or a balance of the two.
- Serve fine-tuned LoRA adapters from one endpoint, on models that support them.
Example endpoint settings
Ready for your security review
Run on our partners' capacity, or inside your own network when your security team requires it. Either way, the deployment fits the controls you already run.
- SOC 2We have completed a SOC 2 audit covering security, availability and confidentiality. How we handle security
- Your security toolsWe routinely add custom integrations for tools such as Netskope and CrowdStrike.
- Air-gapped and restricted networksDeploy air-gapped, on private subnets behind your VPN, or with egress blocked and images served from your own mirror. High-security installations
- Single sign-onConnect your identity provider, including Okta, Microsoft Entra ID and Google Workspace.
Your network
How we think about serving
Questions
How do you set the SLA targets?
We profile your workload, optimize and benchmark the deployment for your model, and agree targets based on what that shows. The targets depend on your request characteristics, concurrency and capacity, so they differ for each deployment.
Do I need my own GPUs?
No. We run your endpoint on dedicated GPUs from our neocloud partners, including new capacity that hasn't reached the open market and isn't shown in our self-service product. There is nothing for you to source or operate. If you need to stay inside your own environment, we can deploy there instead.
Can this run in an air-gapped or restricted network?
Yes. We deploy in air-gapped and other restricted networks, and we add custom integrations for security tools such as Netskope and CrowdStrike. Tell us your requirements early so we can plan the deployment around them.
Can I bring my own model or agent?
Yes. Our catalogs are a starting point. You can bring your own model, your own agents or any open-source agent.
How do I get started?
Talk to us about your model and your traffic. Sign-up for dedicated endpoints is not self-serve, so we set up your organization and your endpoint with you.
I run a GPU cloud. Is this the right page?
Saturn Cloud also sells the platform behind this to GPU cloud operators. See Token Factory and the operator pages.
Bring us your model and your traffic
We will tell you what it takes to profile your workload and what a performance commitment would cover.
Self-service has limited GPU availability. Contact us for more capacity.