A 16 GB Windows PC can run useful local language models, but the right choice is usually a compact model rather than the largest download that appears to fit. Windows, Ollama, the model, the conversation context and any other open applications all compete for the same memory. This guide is based on current official Ollama documentation and model-library records; it is not a benchmark or a claim of personal hands-on testing.
Short answer
Start with Llama 3.2 3B when you want a small general assistant for rewriting, summarising and straightforward instructions. Try Qwen3 4B when multilingual work or a reasoning-oriented response style matters. Qwen3 8B may still fit in 16 GB of system RAM, but its larger model file leaves less headroom and long conversations can increase memory use. For an older or heavily loaded PC, smaller is often more dependable.
Model in Ollama | Listed model size | Practical fit | Main trade-off |
|---|---|---|---|
Llama 3.2 1B | 1.3 GB | Very limited hardware; short rewriting and simple retrieval | Lowest capability of this group |
Llama 3.2 3B | 2.0 GB | General writing, summaries and instruction following | Weak on difficult reasoning and detailed coding |
Qwen3 4B | 2.5 GB | Multilingual prompts and a step up in reasoning | Can be slower than a smaller model |
Qwen3 8B | 5.2 GB | Users willing to trade speed and headroom for more capability | Less room for Windows, context and other applications |
What 16 GB RAM really means
A model’s download size is a useful first filter, not a complete memory requirement. Runtime memory also depends on context length, quantisation, Ollama’s overhead and whether the system uses a CPU, integrated graphics or a supported discrete GPU. A model can launch and still be impractical if Windows begins paging to disk or each response takes too long for the intended task.
Ollama currently requires Windows 10 or later. Before downloading several models, check free storage as well as memory: each model remains on disk until removed. Close memory-heavy browser tabs for the first evaluation and use a short conversation so that model choice, rather than an oversized context, is what you are comparing.
How the main choices differ
Llama 3.2 3B: the sensible starting point
Ollama lists the default Llama 3.2 package at about 2.0 GB and describes the 3B instruction model for tasks including instruction following, summarisation, prompt rewriting and tool use. Those are appropriate expectations for a compact local assistant. It is not a substitute for a leading hosted model on complex analysis, precise factual research or large codebases.
Llama 3.2 1B: when responsiveness matters most
The 1B package is smaller and can be useful for basic rewriting or classification-style prompts. Choose it when the 3B model is uncomfortably slow, not because it will match a larger model. Its answers need particularly careful review when a task has several constraints.
Qwen3 4B: a compact multilingual alternative
Ollama lists Qwen3 variants from 0.6B upward and describes the family as supporting more than 100 languages and dialects. The 4B package is listed at about 2.5 GB, making it a reasonable comparison point on a 16 GB PC. Qwen3 can produce visible reasoning-style output depending on the runner and model settings; that may help inspection but can also make simple answers slower or longer.
Qwen3 8B: possible, but not always comfortable
The listed 8B package is about 5.2 GB. That is below 16 GB, but it does not mean the entire workflow will occupy only 5.2 GB. Test it with the applications you normally keep open. If the system becomes unresponsive or relies heavily on the page file, the 4B model is the better operational choice even when the 8B answer is sometimes stronger.
A fair five-prompt evaluation
Summarise a public document into five points, then check every point against the document.
Rewrite a short email while preserving names, dates and requested actions.
Extract a small set of fields from invented text into a fixed table.
Explain a familiar technical concept at beginner level, then identify omissions yourself.
Ask one prompt in every language you genuinely need and have a fluent reader review it.
Use the same prompt and clean conversation for each model. Record whether the model followed the format, invented facts, omitted constraints or required repeated corrections. Also record time to first response and whether the rest of the PC remained usable. A model that wins a public benchmark can still be the wrong choice for your workload.
Model size and quantisation in plain language
A model name such as 3B or 8B refers broadly to the number of learned parameters, not the amount of RAM it will use. More parameters can improve capability, but they also require more storage, memory and computation. Local runners normally use quantised model files: instead of keeping every model value at high precision, they store a compressed numerical representation. This makes a model much smaller and allows it to run on ordinary hardware, with a possible loss of answer quality.
The package size shown by Ollama is therefore a practical comparison clue. It is not a promise about peak RAM. Ollama also allocates a context cache for the prompt and conversation. Its official FAQ states that the default context is 4,096 tokens and that required memory grows with context length and parallel requests. On a 16 GB PC, leave the context at its default for the first run. Increasing it before there is a real need can make an otherwise suitable model slow or unstable.
Realistic expectations for CPU-only use
Ollama can load a model entirely into system memory when no supported GPU acceleration is available; `ollama ps` reports `100% CPU` in that case. CPU-only inference is useful for private drafts, short summaries, extraction and occasional coding questions, but response generation may be visibly slower than a hosted service or a modern discrete GPU. Long answers, large contexts and 8B-class models demand more waiting. There is no honest universal tokens-per-second figure because processor generation, memory bandwidth, model build, context and background activity all matter.
A dual-core or early low-power laptop processor should start with Llama 3.2 1B or 3B and short outputs. A more recent four-core or six-core desktop can reasonably compare Llama 3.2 3B, Qwen3 4B and Gemma 3 4B. Treat an 8B model as an optional second-stage test. If Windows becomes unresponsive, disk activity remains high or the fan stays at maximum for routine prompts, the model is operationally too large even if it eventually answers.
Installation and first-run workflow
Confirm the PC runs Windows 10 22H2 or newer. Ollama’s current Windows documentation does not support older Windows releases.
Check storage before installation. Ollama states that its Windows installation needs at least 4 GB, in addition to model files that can accumulate into tens of gigabytes.
Download Ollama from its official Windows page and use the normal installer. It installs for the current user and runs in the background.
Open PowerShell and run `ollama run llama3.2`. Ollama downloads the default model and opens an interactive prompt after loading.
Use a short, non-sensitive test prompt. Ask for a fixed format so instruction-following is easy to inspect.
Run `ollama ps` in another terminal to see whether processing is on CPU, GPU or split across both.
Type `/bye` when finished. If memory remains constrained, run `ollama stop llama3.2` to unload the model immediately.
Storage and memory management
On Windows, Ollama stores models under the user profile by default. The official Windows guide documents the `OLLAMA_MODELS` environment variable for moving them to another drive. This is useful when the system SSD is small, but moving models to a very slow hard disk can lengthen load times. Keep enough free space for Windows updates and the page file rather than filling the drive with several similar models.
Only keep models that serve a distinct task. Stop an idle model when another large application needs memory, reduce the context before choosing a smaller quantisation, and avoid parallel requests on a 16 GB machine. Do not disable the Windows page file merely to force an error: paging is slower than RAM, but an automatically managed page file can prevent a sudden application failure.
Troubleshooting common first-run problems
Symptom | Likely cause | Practical response |
|---|---|---|
The command is not recognised | Terminal opened before installation updated PATH | Close and reopen PowerShell; confirm Ollama is running from the Start menu |
Model download fails | Network, proxy or insufficient disk space | Check free space and the official proxy guidance; retry from a stable connection |
The PC freezes or pages heavily | Model, context or other applications exceed comfortable memory | Close heavy applications, stop other models, restore default context or choose a smaller model |
Responses are very slow | CPU-only inference or an oversized model | Check `ollama ps`; shorten output and compare a 1B–4B model |
Answers lose earlier instructions | Context is full or the prompt is unclear | Start a fresh conversation, shorten supplied text and state the required format first |
Model remains in memory | Ollama keeps models loaded briefly for reuse | Use `ollama stop <model>` when the RAM is needed immediately |
Privacy and limitations
Local inference can keep prompt processing on the computer, but privacy still depends on the surrounding workflow. Documents may be copied into chat logs, backups or third-party interfaces. Model output can be inaccurate, biased or insecure, and a local model has no automatic knowledge of current events unless you provide trustworthy material. Do not place credentials in prompts or use an unreviewed answer for medical, legal, financial or other consequential decisions.
Decision: Choose Llama 3.2 3B as the default starting point. Move to Qwen3 4B for multilingual or reasoning-heavy tasks. Try Qwen3 8B only when you can accept slower responses and reduced memory headroom. Drop to a 1B model when keeping the PC responsive matters more than answer depth.
Recommendations by user type
Older laptop, CPU only: start with Llama 3.2 1B, then try 3B only if the PC remains responsive.
General home or office PC: use Llama 3.2 3B for drafting and summaries; compare Qwen3 4B for multilingual prompts.
Developer with a recent CPU but no discrete GPU: compare Qwen3 4B and Gemma 3 4B on a small code-explanation task, while keeping expectations below hosted coding systems.
PC with a supported GPU and 16 GB system RAM: use `ollama ps` to confirm acceleration, but still size the model for available VRAM and system headroom.
Privacy-sensitive user: keep Ollama local-only, do not expose its local API to the network, and verify any third-party interface before granting file access.
Final recommendation
For most 16 GB Windows PCs, a 2–3 GB model package is the practical sweet spot. Begin with one real, low-risk task and judge the result rather than collecting models. Keep the smaller model if it meets the requirement consistently; upgrade only when a documented failure shows that more capability is worth the additional memory and waiting time.
