Tokens, served from the network.
An OpenAI-compatible endpoint for open-weight models, running on SaladCloud's distributed GPUs. Change one base URL and your agents, coding assistants, and apps pay a fraction of frontier-API prices.
Early access. Flat monthly subscription or pay-per-token from your SaladCloud credit balance.
# Works with any OpenAI SDK from openai import OpenAI client = OpenAI( base_url="https://ai.salad.cloud/v1", api_key="your-ai-gateway-api-key", ) r = client.chat.completions.create( model="qwen3.6-35b-a3b", messages=[{"role": "user", "content": "Say hello in one sentence."}], stream=True, )
Open weights, production configs.
Every model supports chat completions, streaming, tool calling, structured outputs, and reasoning, with text and image input.
| Model ID | Best for | Input $/M | Output $/M |
|---|---|---|---|
qwen3.6-35b-a3bQwen3.6 35B-A3B (MoE) | Agentic coding and complex reasoning. Recommended default. | $0.06 | $0.95 |
qwen3.6-27bQwen3.6 27B | General purpose; balances capability and speed. | $0.29 | $1.90 |
qwen3.5-35b-a3bQwen3.5 35B-A3B (MoE) | Cost-efficient agentic tasks and reasoning. | $0.06 | $0.90 |
qwen3.5-9bQwen3.5 9B | High-volume, simple queries; the fastest model in the lineup. | $0.07 | $0.23 |
Per-token rates apply to pay-per-token billing. The flat monthly subscription covers all models with a two-concurrent-request limit. Catalog and prices as published at docs.salad.com; the roster expands as quantized builds meet our quality bar.
The lowest-cost token is the one served on hardware that's already paid for.
Up to 90% lower cost
Frontier-quality open models on consumer and workstation GPUs across the network. No cluster to rent, no reserved capacity to amortize.
Drop-in compatible
Works with Cline, Aider, OpenCode, OpenClaw, Roo Code, Continue, Vercel AI SDK, Goose, and anything else that accepts a custom OpenAI base URL.
Private by default
Prompts and completions are never used for training. Requests are metered, not mined.
Sustainable compute
Every token is served on GPUs that would otherwise sit idle, no new data-center build-out required.
Request a monthly plan, or pay per token.
Flat monthly subscription
One price, every model, unlimited tokens, two concurrent requests. Built for individual developers and agent workflows that run all day.
Request access in the portal →Pay per token
Usage billed against your organization's SaladCloud credit balance at the published per-million rates. No concurrency cap; scales with your traffic.
Pay-per-token onboarding →Need your own models, or your own network?
Bring your own model
Deploy vLLM, TGI, or Ollama recipes on Salad Container Engine for a flat hourly GPU rate. Fine-tunes, custom quantizations, any open weights.
See recipes →Reserved capacity
Pin inference to dedicated RTX PRO 6000 Blackwell servers with 96 GB of VRAM for larger models and longer contexts.
Salad Dedicated →Private endpoint
The same gateway, deployed inside your organization on GPUs you already own. Data never leaves your network.
Private Cloud →Questions engineers ask first.
Is it really OpenAI-compatible?
Yes. Set OPENAI_BASE_URL=https://ai.salad.cloud/v1 and use your AI Gateway key as the API key. Chat completions, streaming, tool calling, and structured outputs behave as the OpenAI SDKs expect.
Where do the GPUs come from?
Models run on SaladCloud's distributed network, primarily RTX 5090 and 4090 class hardware today. Each model is served with the context length and quantization published on its model card; we do not silently downgrade either.
What about latency?
The network is globally distributed, often on residential connections, so time-to-first-token is higher than a co-located endpoint. For interactive coding assistants and agents it is well within range; for sub-100-ms voice loops, ask us about Dedicated.
Can I get a different model?
The catalog grows as quantized builds pass our accuracy checks. Tell sales which model and context length you need; if there's demand we prioritize it, and you can always deploy your own on Container Engine today.