The “B” in a model name means billions of parameters. Qwen 3.5 4B has about 4 billion, Qwen 3.5 9B about 9 billion. Parameters are the numbers a model learned during training, and more of them usually means more knowledge and better reasoning, at the cost of more memory and slower answers. On a phone, the parameter count is the first thing that tells you whether a model will fit.
The other letters in model names (E2B, A3B, 128K, Q4_K_M) mean different things, and mixing them up is how people end up downloading something their phone can’t run.
What is a parameter? #
A language model is a very large set of numbers arranged in layers. When you type, your words pass through those layers, and the numbers decide which word is most likely to come next. Those numbers are the parameters, sometimes called weights.
Training sets their values by showing the model enormous amounts of text. A model with more parameters has more room to store patterns: facts, grammar, styles of reasoning. That’s why a 9B model generally knows more and makes fewer mistakes than a 0.8B model from the same family.
How much storage and RAM does each size need? #
The file size depends on two things: how many parameters there are and how many bits each one is stored with. Phone models are “quantized,” meaning each parameter is squeezed down from 16 bits to around 4 to 8 bits. That’s what makes them fit.
Here are real numbers from the Personal LLM catalog, which lists the download size and the minimum RAM for each model:
| Model | Size label | Download | Needs RAM |
|---|---|---|---|
| Qwen 3.5 0.8B | 0.8B | 0.81 GB | 2 GB+ |
| Gemma 4 E2B | 2.3B effective | 2.04 GB | 3 GB+ |
| Ministral 3 3B | 3B | 2.15 GB | 3 GB+ |
| Qwen 3.5 4B | 4B | 2.74 GB | 3 GB+ |
| Gemma 4 E4B | 4.5B effective | 3.00 GB | 5 GB+ |
| Qwen 3.5 9B | 9B | 5.68 GB | 6 GB+ |
| GLM 4.6V Flash | 9B | 6.17 GB | 8 GB+ |
A useful rule of thumb: at the 4-bit quantization most phone models use, budget a bit over half a gigabyte per billion parameters, plus room for the conversation. The app reads your phone’s RAM and shows a “Fits your device” badge, so you don’t have to do the math.
What does the “E” in Gemma 4 E2B mean? #
E stands for “effective.” Google describes Gemma 4 E2B and E4B as effective-parameter models built for phones, edge devices and browsers. Part of these models’ parameters can be stored and loaded more cheaply than the rest, so the model runs with roughly the compute of a 2B or 4B model even though its total parameter count is higher.
In practice: judge E2B like a small model for speed, and expect somewhat better quality than a plain 2B model. The catalog lists E2B at 2.3 billion effective parameters and E4B at 4.5 billion.
What does “A3B” or “A17B” mean? #
That’s a mixture-of-experts (MoE) model. Qwen 3.5 comes in sizes like 35B-A3B and 397B-A17B, according to Qwen’s release notes. The first number is the total parameter count. The “A” number is how many are active for each word.
An MoE model is split into many “experts,” and a small router picks a few for each token. So a 35B-A3B model computes about as fast as a 3B model, but it still has to keep all 35 billion parameters in memory. That’s why MoE models are great on computers with lots of RAM and not a good fit for phones. The phone-sized Qwen 3.5 models (0.8B, 2B, 4B and 9B) are regular dense models.
What do 128K and 256K mean? #
Those numbers are the context window: how many tokens (word pieces) the model can consider at once. They’re not parameters. Ministral 3 3B supports 256K tokens, Gemma 4’s small models 128K, and Qwen 3.5 lists 262,144 tokens natively.
On a phone you won’t use anything close to that. The context uses memory too, and the more you use, the slower each reply gets. Personal LLM starts at 2,048 tokens and caps vision models at 4,096 to avoid running out of memory. You can raise it in settings if your phone has RAM to spare.
What do Q4_K_M and Q8_0 mean? #
Those are quantization labels in GGUF files, the format phone apps like Personal LLM load. The number is roughly the bits per parameter: Q4 is about 4 bits, Q8 about 8. Lower is smaller and faster, higher is closer to the original quality. Q4_K_M is the common sweet spot for phones. Our guide to GGUF quantization explains every label.
What do “instruct,” “it” and “base” mean? #
- Base models are trained only to continue text. They’re not good at following instructions.
- Instruct or it (instruction-tuned) models are trained further to follow requests and chat. These are what you want in a chat app.
- Thinking or reasoning variants write out step-by-step working before answering. Qwen 3.5 models think by default, and Personal LLM shows that working in a collapsible panel when you turn thinking mode on.
Does a bigger model always give better answers? #
Usually within the same family, not always across families.
- Newer beats older. A 4B model from 2026 can outperform a 7B model from two years earlier, because training methods and data improved.
- Task matters. For rewriting a paragraph you pasted, 4B and 9B often give similar results. For tricky reasoning or obscure facts, the 9B pulls ahead.
- Thinking mode helps small models. Letting a 4B model reason step by step can close part of the gap on math and logic, at the cost of speed.
- Speed is part of quality. A 9B model that crawls on your phone may be less useful than a 4B that answers at reading speed.
If you want to measure the speed side on your own phone, see how fast a local LLM runs on a phone.
Which size should you pick for your phone? #
| Phone RAM | Start with | Try next |
|---|---|---|
| 3 to 4 GB | Qwen 3.5 4B or Gemma 4 E2B | Qwen 3.5 0.8B if it’s slow |
| 6 GB | Qwen 3.5 4B | Gemma 4 E4B, or Qwen 3.5 9B for harder questions |
| 8 GB or more | Qwen 3.5 9B | GLM 4.6V Flash for photos and documents |
More detail by phone is in how much RAM you need to run an LLM.
Frequently asked questions #
What does 7B mean in an AI model? #
It means the model has about 7 billion parameters, the learned numbers that make up the model. The same “B” convention applies to 4B, 9B, 70B and so on.
How many parameters does ChatGPT have? #
OpenAI doesn’t publish parameter counts for its current models. They are widely understood to be far larger than anything a phone can run, and they run on data-center hardware.
Can a phone run a 70B model? #
No. Even at 4-bit quantization a 70B model needs roughly 40 GB of memory, several times what phones have. Phone-friendly models top out around 9B today.
Is Qwen 3.5 9B better than Qwen 3.5 4B? #
Generally yes: it knows more and reasons better. It’s also twice the download (5.68 GB vs 2.74 GB), needs at least 6 GB of RAM, and is slower. Many people use the 4B for everyday chat and switch to the 9B mid-conversation for harder questions.
Why is the file smaller than the parameter count suggests? #
Because of quantization. Stored at 16 bits, a 4B model would be about 8 GB. At around 4 bits per parameter it shrinks to under 3 GB, with a small loss in quality.