ChatGPT, 
Claude, 
Gemini, 
Grok — all four for 
If you've installed Ollama and started browsing models, you've probably noticed names like "7B," "13B," or "70B" attached to them. These numbers aren't random, and they matter a lot for how well a model will run on your server. This article explains what they mean and how to use them to pick a model that fits your plan.
What Does the "B" Actually Mean?
The "B" stands for billions, as in billions of parameters. A parameter is roughly one small piece of what the model learned during its training, similar to one tiny adjustment out of millions that together shape how it understands and responds to language. A "7B" model has 7 billion of these; a "70B" model has ten times as many.
As a simple rule: more parameters generally means a more capable model, better at understanding nuance, following complex instructions, and handling harder tasks. But more parameters also means the model needs more memory to run and takes more computing power to produce each response. There's a real tradeoff between capability and how much hardware you need.
Common Model Sizes and What They Need
Here's roughly what to expect from each size range, based on Ollama's own recommendations using its default settings, along with which Self-Managed VPS or VDS plan comfortably covers it (with some headroom, not just the bare minimum):
| Model Size | Minimum RAM Needed | General Use Case | Recommended Plan |
|---|---|---|---|
| Under 3B | 4GB or less | Very simple tasks, quick replies, basic testing | NVMe 4 (2 vCPU, 4GB RAM) |
| 7B | At least 8GB | General chatting, everyday questions, light writing help | NVMe 16 (8 vCPU, 16GB RAM), or Dedicated 2 (2 vCPU, 12GB RAM) for guaranteed resources |
| 13B | At least 16GB | More nuanced writing, more reliable answers, moderate coding help | Dedicated 4 (4 vCPU, 24GB RAM); NVMe 16 works but leaves little headroom |
| 70B | At least 64GB | Complex reasoning, advanced coding help, demanding business use | Dedicated 16 (16 vCPU, 96GB RAM) |
Plan names and specs may change — check the Self-Managed VPS and VDS pages for current plans and specs.
These are minimums, not comfortable operating room. Running right at the minimum often means a slower response and little headroom for anything else running on your server at the same time.
DeepSeek Comes in These Exact Sizes
DeepSeek, one of the one-click AI apps available on Bluehost, is a good real-world example of this. It isn't a single fixed model, it's released in several sizes: 1.5B, 7B, 8B, 14B, 32B, 70B, and a full-scale 671B version. When you install DeepSeek, you're choosing which of these to run, and that choice determines both how capable it is and how much RAM you'll need.
The smaller versions (1.5B up to around 14B) are realistic to run on a Self-Managed VPS. The 32B and 70B versions need considerably more RAM, closer to what a VDS plan offers. The full 671B version is in a different league entirely, it's designed for large multi-server setups and isn't practical to run on a single VPS or VDS plan. For almost everyone self-hosting DeepSeek, one of the smaller, distilled versions is the realistic choice.
If a model's minimum requirement is close to your total plan RAM, it's worth sizing up. Your operating system, and any other apps you're running, need some of that RAM too, not just the model itself.
This Applies to More Than Just Chat Apps
It's easy to assume model sizing only matters for a chat setup like Open WebUI, but it actually applies to any AI app that ends up connected to a local model through Ollama. Several of Bluehost's agentic AI apps genuinely support this: Claude Code, OpenClaw, Hermes Agent, n8n, Dify, Paperclip, and Sim can each be pointed at a local model instead of a cloud AI provider, and once you do, the same RAM requirements from the tables above apply, whether that model is 7B, 13B, or 70B.
A couple of apps work a little differently. Nanoclaw is built specifically on the Claude Agent SDK, so it can hand off individual subtasks, like summarizing text, to a local model, but its main orchestrator still relies on a cloud AI provider rather than being fully replaceable with one. BMAD isn't a model host at all: it's a workflow layer that installs on top of an AI coding tool you already have, such as Claude Code. Whether BMAD ends up using a local model comes down entirely to whichever underlying tool you've connected it to. GatorClaw is a bit different again: it's built on the same underlying ecosystem as OpenClaw, but as the simplified, guided version aimed at non-technical users, and Bluehost's own documentation describes it as relying on third-party LLM providers like OpenAI, Anthropic, or Google, with no mention of a local Ollama option.
In every case, the app itself is usually lightweight; it's the model doing the actual thinking behind it that determines how much RAM you really need.
A Quick Note on Compressed Models
You may also notice tags on model names, like q4 or q8. This refers to quantization, a way of compressing a model so it takes up less memory and runs faster, in exchange for a very small dip in accuracy. Ollama uses a commonly recommended level of compression by default, which is why the RAM figures above are often lower than you might expect from a model's full, uncompressed size. If you're tight on RAM, choosing a more compressed version of a model is a reasonable way to make it fit.
Summary
The number attached to a model's name (7B, 13B, 70B, and so on) tells you how many billions of parameters it has, which is a rough measure of both its capability and how much RAM it needs to run. Bigger generally means smarter, but also heavier. Check your Self-Managed VPS or VDS plan's RAM against the model's minimum requirement before installing it, and leave some room to spare rather than running right at the limit.