How to run an AI model on your own computer

1. Check your computer

On Mac, LM Studio documents Apple Silicon M1/M2/M3/M4 and macOS 14 or newer. Intel Macs are unsupported. The app maker suggests 16 GB or more RAM, while noting smaller models and modest context sizes may be usable on 8 GB Macs. On Windows, it documents x64 with AVX2, or ARM on Snapdragon X Elite; it suggests at least 16 GB RAM and 4 GB dedicated GPU memory. These are app-level requirements and suggestions, not verification of this model.

Check your chip, operating system and installed RAM in the computer’s system information. Installed memory differs from memory currently available after the operating system and other apps use some. Keep system RAM and dedicated GPU memory separate. Apple unified memory is one shared pool. Context length, runtime, format and load settings change memory use. Bright does not know this candidate’s speed, runtime memory or output quality on your machine.

Sources: LM Studio: System Requirements · LM Studio: Get started

2. Install desktop LM Studio

Open the official download page and choose the installer for your operating system and processor under “Download LM Studio”. The page also offers Bionic and the headless llmster daemon; this guide follows the desktop chat app. Follow the installer’s instructions and open LM Studio. Record the app version you installed so you can identify it if something changes.

Sources: Download LM Studio · LM Studio: Get started

3. Find the exact model

Open Discover and search for ggml-org/SmolLM3-3B-GGUF, or paste its Hugging Face repository URL. LM Studio documents both search methods. Confirm the publisher is ggml-org and the filename is SmolLM3-Q4_K_M.gguf. This is a quantized conversion of HuggingFaceTB/SmolLM3-3B, not the original full-precision checkpoint. GGUF is the file format; Q4_K_M identifies this quantization.

Sources: LM Studio: Download an LLM · ggml-org: SmolLM3-3B-GGUF files · HuggingFaceTB: SmolLM3-3B model card

4. Download the Q4_K_M file

Select that exact file’s download option. The checked repository lists it at 1.92 GB and identifies Apache 2.0 licensing, also stated by the original model card. Read the linked license before reuse. Allow disk space for the file, the app and any needed runtime. A completed download tells you that you have the artifact; loading and generating are separate steps.

Sources: ggml-org: SmolLM3-3B-GGUF files · HuggingFaceTB: SmolLM3-3B model card · LM Studio: Download an LLM · LM Studio: Offline Operation

5. Load it and send a first prompt

In Chat, open the model loader and select the downloaded SmolLM3 model. Review the load settings, including context length, then load it. Start a new chat and ask: “Rewrite this sentence more clearly: We moved the meeting to Sunday at one.” If loading fails, keep the error and app version rather than interpreting the file size as a compatibility test.

Sources: LM Studio: Get started · LM Studio: Manage chats

6. Check a reply with the picnic exercise

Try Bright’s fictional exercise: “Write a short message to the picnic group using only these fictional notes. Preserve the changed day, time and place. Include who brings what. Flag undecided items rather than inventing a plan. Notes: The picnic moved from Saturday to Sunday. It starts at 1 p.m. at River Park. Lee brings fruit. Sam brings sandwiches. The rain plan is undecided.”

Check the reply against five facts: Sunday; 1 p.m.; River Park; Lee brings fruit and Sam brings sandwiches; rain remains undecided. Mark any omission or invented detail. The exercise is a reader check, not a benchmark or a Bright test result. The maker cautions that generated content can be inaccurate or inconsistent; verify important information yourself.

Sources: LM Studio: Manage chats · HuggingFaceTB: SmolLM3-3B model card

7. Try the downloaded model offline

Finish the model and required runtime downloads first. Then disconnect your network and send another fictional prompt in the local Chat tab. LM Studio documents local chatting without internet connectivity. Model search, new downloads, runtime downloads and update checks need connectivity. Keep this exercise in local chat without connected tools; external integrations have their own data paths. This is a check you can perform, not an offline or privacy test Bright has performed.

Sources: LM Studio: Offline Operation

8. Remove the model to reclaim disk space

Open My Models and identify the exact SmolLM3-Q4_K_M.gguf download. The official download documentation points to My Models for the models directory, and the import documentation describes publisher/model/file folders. The official pages checked here do not establish the current delete-button wording. If your version offers a removal action, read its confirmation carefully.

Alternatively, quit the app and use Finder on Mac or File Explorer on Windows to open the models directory shown by your installation. Remove only the identified GGUF file using your operating system’s normal file deletion. Reclaiming space may require emptying the Trash or Recycle Bin. This file-removal step is Bright’s practical inference from the documented directory structure, not a documented LM Studio button sequence. Deleting the weights does not delete your saved chats.

Sources: LM Studio: Download an LLM · LM Studio: Import Models · LM Studio: Manage chats

What this guide does not establish

• This guide is documentation-led and untested by Bright on Mac or Windows. It reports no hands-on benchmark, generated model reply or measured compatibility.
• Documentation and interface labels may change. Recheck this guide when LM Studio changes major version or its requirements change.
• The worked example is a documented candidate. The listed file size does not establish total runtime memory, supported context length, speed or quality.
• Model licensing, conversion provenance and app terms are separate. Apache 2.0 for the model does not license the LM Studio app or third-party documentation.

Primary sources

Source-reviewed 2026-10-11. Source review means Bright compared this explanation with the linked records; it is not independent replication or expert peer review by Bright.

Continue into the evidence