Use Ollama on Rollin VPS if:
You run chat or RAG with models up to 13B in Portuguese, predictable volume. Critical privacy. Want minimum latency for Brazilian users.
Technical comparison
Ollama, Together.ai or RunPod: which to choose for open-source LLM?
Hosting Llama 3, Mistral, Qwen and other open-source models became common strategy in 2026 — be it for cost, privacy or customization. The three main options: run Ollama on VPS/CPU, use a managed API like Together.ai, or rent GPU on-demand at RunPod.
TL;DR
Ollama on Rollin VPS (CPU) runs small quantized models (Llama 3 8B, Phi-3, Qwen 2.5 7B) with 30-80ms latency for Brazil and fixed cost of R$ 89.90-199.90/month. Together.ai is a managed API with fast inference and per-token pricing, no ops. RunPod offers GPU on-demand for large models (70B+) or fine-tuning.
Comparison table
| Feature | Ollama on Rollin VPS | Together.ai | RunPod |
|---|---|---|---|
| Service type | Self-hosted CPU | Managed API (serverless) | GPU on-demand (IaaS) |
| Hardware | AMD EPYC + NVMe (CPU) | Managed GPUs | RTX 4090 / A100 / H100 |
| Supported models | Llama 3 8B, Mistral 7B, Qwen 2.5, Phi-3 | Llama 3 70B, Mixtral, DeepSeek, +100 | Any open-source model |
| Brazil latency | 30-80ms (SP datacenter) | 200-400ms (US/EU) | 150-350ms (varies) |
| Entry cost | R$ 89.90/mo (Pro 10) | Pay-per-token | US$ 0.30-3.50/hr GPU |
| Cost per million tokens | Amortized in fixed | ~US$ 0.20-0.90 | Calculated per GPU hour |
| Large 70B+ models | Unfeasible on CPU | Natively supported | Yes, adequate GPU |
| Fine-tuning | Limited (slow CPU) | Yes, managed | Yes, full GPU control |
| Privacy (where is data?) | Your VPS in Brazil (LGPD) | Together servers (US, +DPA) | RunPod servers (US/EU) |
| Cold start | Zero | ~1-5s (serverless) | 30-120s (GPU boot) |
| Throughput tokens/sec | 20-60 tok/s (8B on CPU) | 50-200 tok/s | 100-500 tok/s |
| Billing | Fixed in BRL (R$) | Per use in USD | Per hour in USD |
| Vendor lock-in | Zero (open-source) | Medium (proprietary API) | Low |
| Operation | You manage Ollama | Zero — just API | You set up Docker + container |
| Portuguese human support | Yes, 24/7 via Rollin | English only | English only |
Ollama on Rollin VPS Pros
- Predictable fixed BRL cost
- Minimum latency for Brazil
- Data never leaves your VPS
- Runs on same machine as n8n, EvolutionAPI, Qdrant
- Zero cold start
- 24/7 Portuguese human support
- Open-source: Llama, Mistral, Qwen, Phi
Together.ai Pros
- Immediate access to 100+ models
- Fast inference with managed GPU
- Pay-per-token
- No ops
- Excellent for variable loads
- Good documentation and SDKs
- Supports managed fine-tuning
RunPod Pros
- GPU on-demand: RTX 4090, A100, H100
- Flexible hourly pricing
- Runs any model
- Ideal for fine-tuning
- Serverless mode also available
- Active community, Docker templates
Ollama on Rollin VPS Cons
- CPU limits large models
- Lower throughput than GPU
- No auto-scaling
- Serious fine-tuning needs GPU
- You manage Ollama updates
Together.ai Cons
- 200-400ms latency from Brazil
- USD billing with IOF
- Data on Together servers — DPA for LGPD
- At high stable volume, more expensive than self-hosted
- No Portuguese support
- Rate limits on popular models
RunPod Cons
- 30-120s cold start
- USD hourly billing
- You are responsible for Docker
- Variable GPU availability
- No Portuguese support
- No BR region
When to choose each
Use Together.ai if:
You need Llama 3 70B without investing in GPU. Variable loads — prototypes, peaks. Small team without DevOps.
Use RunPod if:
You will do fine-tuning. Need GPU for specific workloads. Want full environment control.
Use combination if:
Ollama on Rollin VPS for production + Together.ai for fallback on large models + RunPod for fine-tuning.
Verdict
For most Brazilian cases, Ollama on Rollin VPS delivers the best cost-benefit with total privacy and minimum latency. Honestly, if you need 70B+ model, Together.ai is clearly superior. RunPod is the right tool for fine-tuning. Rollin Host does not offer dedicated GPU in 2026, so if your case is serious fine-tuning, use RunPod without guilt.
Frequently asked questions
Can I run Llama 3 8B on CPU?
Yes. With Q4 or Q5 quantization (GGUF), Llama 3 8B runs on a VPS with 8-16 GB RAM and AMD EPYC delivers 20-40 tokens/second.
How much does Together.ai cost in 2026?
Together.ai charges per token. Llama 3 8B costs ~US$ 0.20/M tokens, Llama 3 70B around US$ 0.90/M tokens.
Does RunPod have a Brazilian datacenter?
In 2026, RunPod has no Brazilian region. Most-used regions are US-East, US-West and EU.
Does Ollama support function calling?
Yes, since version 0.3+ Ollama supports tool/function calling with compatible models.
Can I fine-tune on Ollama?
Technically yes, but impractical on CPU. For serious fine-tuning, use RunPod with GPU.
Is Together.ai LGPD compliant?
Together.ai offers signable DPA. Since data passes through US servers, review the case with your DPO.
Which Rollin VPS for Ollama?
Pro 10 (R$ 89.90/month) runs Llama 3 8B Q4. For Mistral 7B + simultaneous RAG, Pro 20.
What is the throughput of Llama 3 70B on RunPod?
On an A100 80GB, Llama 3 70B FP8 delivers ~80-150 tokens/second. On H100, rises to ~200-400 tok/s.
Can I embed with Ollama?
Yes. Ollama supports embedding models like nomic-embed-text and mxbai-embed-large.
Does Together.ai have chat playground?
Yes, Together.ai has web playground to test models before integrating via API.
How to set up Ollama on a Rollin VPS?
curl -fsSL https://ollama.com/install.sh | sh — then ollama pull llama3.1:8b-instruct-q4_K_M.
RunPod vs Vast.ai, which is best?
RunPod has more polished UX and official templates. Vast.ai is cheaper but with more operational friction. For teams without DevOps, RunPod.
Ready to host your open-source LLM with privacy?
VPS Cloud AMD EPYC + NVMe from R$ 89.90/month. Ollama, Llama 3, Mistral in minutes.
See VPS for LLM