Short answer: yes, an old Windows PC with 8GB of RAM and no big graphics card can run a useful AI for free. Install Ollama, then type ollama run lfm3. LFM3 1.7B is a small text model from Liquid AI (released Oct 27, 2026). Its default Ollama download is about 1.1GB, Liquid says it takes about 1GB of memory at 4-bit, and it has a 256K token context window and tool calling. It is good for short emails, summaries, rewrites and simple questions. It is not built for coding or hard multi step reasoning. Speed depends on your CPU.
Every spec below comes from Ollama's LFM3 page, Liquid AI's LFM3 announcement and Liquid AI's LFM3 1.7B model card, checked Nov 10, 2026. We did not benchmark it on a test PC, so there are no speed numbers here, only what to expect. No prices either. Click through for the current price.
Why an old PC is worth a try right now
RAM and graphics cards have gotten a lot more expensive in recent months. So before you buy anything, it makes sense to see what the PC you already own can do. Small models like LFM3 are made for exactly this: phones, laptops and ordinary desktops, not expensive AI workstations. If you are not sure what your PC has, run our free Check My PC tool first.
What LFM3 is (verified specs)
| Spec | LFM3 1.7B (text) | Source |
|---|---|---|
| Maker | Liquid AI. Announced Oct 27, 2026 as the first release of the LFM3 family | Liquid AI |
| Parameters | 1.73 billion | Liquid AI / Ollama |
| Memory at 4-bit | About 1GB ("fits on any phone or laptop") | Liquid AI / Ollama |
| Context window | 256K tokens (input plus output) | Liquid AI / Ollama |
| Design | Tri-hybrid: 24 layers, 13 gated short convolutions, 5 sliding window attention, 6 global attention | Liquid AI |
| Built for | Local agents, tool calling, structured outputs, fine tuning for one job | Liquid AI |
| Not recommended for | Agentic coding or complex multi step reasoning (its small size) | Liquid AI |
| Languages | English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Hindi, Russian, Polish | Liquid AI |
| License | LFM Open License v1.0 (open weights) | Liquid AI |
| Family | Base, text, vision, audio, omni and decision models share one backbone. Ollama's lfm3 page lists the text model (text input) | Liquid AI / Ollama |
Which LFM3 download to pick in Ollama
Ollama lists these tags for lfm3 (sizes as shown on the Ollama tags page, Nov 10, 2026). On an old 8GB Windows PC, just use the default.
| Tag | Download size | Use it when |
|---|---|---|
lfm3 or lfm3:1.7b (default) | 1.1GB (listed as 1.1GB to 1.7GB) | Start here. Best fit for 8GB PCs |
lfm3:1.7b-q4_K_M | 1.1GB | Same 4-bit size, pinned by name |
lfm3:1.7b-qad-q4_0 | 1.0GB | The smallest tag, for the tightest PCs |
lfm3:1.7b-q8_0 | 1.8GB | A bit more accurate, needs more memory |
lfm3:1.7b-bf16 | 3.5GB | Full precision. Too heavy for most 8GB PCs |
-mlx-bf16, -mxfp8, -nvfp4 | 3.5GB / 2.1GB / 1.7GB | Marked MLX on Ollama. Skip these on an old Windows PC |
Will it fit in 8GB?
Yes, with room to spare for the model itself. Windows and your browser already use a good share of 8GB, so the about 1GB model plus its working memory is what makes LFM3 a comfortable fit where a 7B or 8B model (around 5GB) would be tight. Two things to know:
- Context costs memory. LFM3 can take up to 256K tokens, but Ollama starts at 4K tokens on machines with under 24GB of graphics memory (see Ollama's context length page). Raising it uses more RAM. On 8GB, keep the default and paste in only the part you need.
- Close the extras. Every open browser tab takes memory the model could use. If the PC starts swapping to disk, everything slows down.
For the full math on memory, see how much RAM and VRAM Ollama needs.
Set it up in 5 steps
- Check your PC. Task Manager > Performance shows your RAM and whether you have a graphics card. Ollama needs Windows 10 22H2 or newer, Home or Pro, and at least 4GB of free disk space for the program, plus room for models (Ollama's Windows page). The installer does not need administrator rights.
- Install or update Ollama. LFM3 is brand new, so an old copy of Ollama may not know it. Get the latest installer from ollama.com. Our Windows install guide walks through it.
- Download and chat. Open PowerShell or Command Prompt and type
ollama run lfm3. The first run downloads about 1.1GB, then a chat prompt appears. - See where it runs. In a second window, type
ollama ps. It shows how much memory the model uses and whether it is on the CPU or a GPU. - Keep answers steady. Liquid AI's model card suggests low randomness settings (temperature 0.1, top_k 50, repetition penalty 1.1). Small models do better with clear, short instructions and the text pasted in.
What to expect for speed (honest version)
We have not timed LFM3 on a real old Windows PC, so we will not give you a tokens per second number. Here is what is safe to say:
- Without a graphics card, it runs on the CPU. Answers type out word by word. A newer CPU with more cores and faster memory is quicker than an older one.
- Small models are the ones that stay usable on CPU. Liquid AI says it tuned LFM3's design on ordinary laptop and phone CPUs, not just big GPUs. A 1.7B model has far less to work through per word than an 8B model.
- Short prompts feel faster. Long pasted documents take a while to read in before the first word appears.
- A small graphics card can still help. If your PC has an NVIDIA card (driver 551.61 or newer) or a supported AMD Radeon card, Ollama will try to use it. Check with
ollama ps.
Good jobs for LFM3 on an old PC
- Draft a short email, note or letter, then edit it yourself.
- Summarize or rewrite text you paste in (a long email, a manual page, a recipe).
- Turn messy notes into a clean list, checklist or simple form (structured output is a listed strength).
- Simple questions and brainstorming, privately. Nothing leaves your PC once the model is downloaded (see Is AI spying on me?).
Where it falls short
- Facts and trivia. A 1.7B model does not hold much world knowledge. Liquid AI says to use retrieval or search for knowledge heavy tasks. Double check anything important.
- Coding and hard reasoning. Liquid AI says it is not recommended for agentic coding or complex multi step reasoning.
- Pictures. Ollama's
lfm3is text only. For reading a photo, see the backup list below.
Backup small models that also fit 8GB
If LFM3 does not suit you, these small models are on Ollama too (default download sizes from each Ollama tags page, Nov 10, 2026). Try one, keep the one you like, and delete the rest with ollama rm.
| Model | Ollama command | Download | Notes |
|---|---|---|---|
| Gemma 3 1B (Google) | ollama run gemma3:1b | 806MB | The lightest option here |
| Llama 3.2 1B (Meta) | ollama run llama3.2:1b | 1.3GB | Tiny and simple |
| Qwen 3 1.7B (Alibaba) | ollama run qwen3:1.7b | 1.4GB | Same size class as LFM3 |
| Llama 3.2 3B (Meta) | ollama run llama3.2:3b | 2.0GB | A classic old PC pick |
| Qwen 3 4B (Alibaba) | ollama run qwen3:4b | 2.5GB | Bigger, a bit smarter, slower |
| Gemma 3 4B (Google) | ollama run gemma3:4b | 3.4GB | Can also look at images. Tight on 8GB |
New to model names? What is Ollama? The AI models explained covers what the "b" numbers mean in plain words.
When an upgrade is worth it
- 16GB of RAM is the cheapest big step. It opens up 7B to 8B models for better writing and answers. See the best Ollama models for 16GB RAM. Check DDR3 vs DDR4 and desktop vs laptop size before you buy (Task Manager > Performance > Memory shows the form factor).
- An SSD helps any old PC that still boots from a hard drive. Models and Windows load much faster.
- A 12GB graphics card is the next step if you want fast answers. See the best budget GPU for Ollama, or plan a full budget AI PC.
More guides live on the Local AI hub.
FAQ
Can a PC with 8GB of RAM run AI?
Yes, if you pick a small model. LFM3 1.7B downloads at about 1.1GB in Ollama's default tag, and Liquid AI says it needs about 1GB of memory at 4-bit precision. That leaves room for Windows on an 8GB PC. Close extra browser tabs while it runs.
Do I need a graphics card to run LFM3?
No. Ollama runs models on the CPU when there is no supported graphics card. Replies come out slower than on a GPU, and how fast depends on your processor, but a small 1.7B model is the kind of model that stays usable on CPU alone.
What is LFM3 good at?
Liquid AI built LFM3 1.7B for local agents, tool calling, structured outputs like lists and forms, and fine tuning for one job. In everyday terms: drafting short emails, rewriting or summarizing text you paste in, and simple questions. Liquid says to pair it with search or your own documents for fact heavy questions.
What should I not use LFM3 for?
Liquid AI says it is not recommended for agentic coding or complex multi step reasoning because of its small size. For coding or harder problems, a 7B to 8B model on 16GB of RAM, or a PC with a 12GB graphics card, is a better fit.
Does LFM3 really have a 256K context window?
Yes, Liquid AI and Ollama both list 256K tokens. But Ollama starts with a smaller default (4K tokens on machines with under 24GB of graphics memory) and a longer context uses more memory. On an 8GB PC, keep the default and paste in only what you need.