You can install llmfit with Scoop on Windows, Homebrew on macOS or Linux, or uv on any supported Python platform. Run llmfit, let it detect your RAM, CPU and supported GPU, then filter its model list by fit, use case, runtime and licence.
The useful part is not simply getting a high score. llmfit shows the quantisation, estimated memory use, context and likely run mode behind each recommendation. You can then download a model through a supported local runtime and replace the estimate with a benchmark from your own machine.
Documentation-verified scope: I prepared this guide on 18 August 2026 from the llmfit 1.1.10 release, repository, TUI guide, provider documentation and platform-support notes. I did not install the program or benchmark a model. The steps below are documentation-verified, not a claim of hands-on testing.
Quick answer: install and run llmfit
On Windows with Scoop:
scoop install llmfit
llmfit doctor
llmfitOn macOS or Linux with the project’s Homebrew tap:
brew install AlexsJones/llmfit/llmfit
llmfit doctor
llmfitIf you already use uv:
uv tool install -U llmfit
llmfit doctor
llmfitMy recommendation: install llmfit first and use it to shortlist two or three sensible models. Do not download the top-scoring model blindly. Inspect the suggested quantisation and context, confirm that your chosen runtime supports it, then measure it on the workload you actually care about.
What llmfit does
llmfit is an MIT-licensed terminal tool for answering a common local-AI question: which language models are realistic on this computer? It detects hardware and scores models across memory fit, estimated speed, quality and usable context.
It also understands several local runtimes, including Ollama, llama.cpp, MLX, Docker Model Runner and LM Studio. If a compatible runtime is available, llmfit can identify installed models and, for supported providers, help download a selected model.
llmfit is not an inference engine. It does not make an oversized model fit, create extra VRAM or guarantee the displayed speed. Its initial figures may come from formulas or community measurements. The final test is still loading the exact quantisation in your runtime and running a representative prompt.
Version 1.1.10 was released on 17 August 2026. It added the Qwen3.8 family and RamaLama discovery, and corrected several model and provider details. That matters because a stale model catalogue can produce a tidy-looking recommendation for the wrong architecture or file.
What you need
- A Windows, macOS or Linux computer.
- A terminal with permission to install a user-level program.
- Scoop on Windows, Homebrew on macOS/Linux, or
uvas an alternative. - A local inference runtime if you want llmfit to download, serve or benchmark a model. Ollama is the easiest starting point; llama.cpp offers more direct GGUF control.
- Enough free storage for the model you choose. A model fitting in memory does not mean its files are small.
Hardware detection varies by platform. Linux has the broadest documented GPU detection. Apple Silicon is fully supported through unified-memory detection. On Windows, llmfit currently documents RAM and CPU detection plus NVIDIA detection through nvidia-smi. If you use an AMD or Intel GPU on Windows, expect to verify or override the memory value manually.
Step 1: install llmfit
Windows
Open PowerShell and install the current package through Scoop:
scoop install llmfitThe project states that its Windows release binaries are signed through SignPath. Scoop also gives you a cleaner update and uninstall path than downloading an arbitrary executable from a mirror.
If Scoop is not installed, use Scoop’s official setup instructions. Review what the package manager will add before approving it.
macOS or Linux
The project recommends its Homebrew tap for a prebuilt binary:
brew install AlexsJones/llmfit/llmfitA Homebrew Core formula is also available:
brew install llmfitThe project-specific tap may reach a new release before the core formula updates. Check llmfit --version after installation if a newly added model family is important to you.
Install with uv
If uv is already part of your Python tooling, install llmfit as an isolated command:
uv tool install -U llmfitYou can also try it without keeping a permanent installation:
uvx llmfitThe repository offers a curl-based installer as well, but a package manager is easier to audit, update and remove.
Step 2: verify hardware detection
Before trusting the model table, run:
llmfit doctorCompare the reported RAM, CPU, GPU and VRAM with your operating system. On an NVIDIA system, nvidia-smi should show the same GPU and a similar dedicated-memory total. On Apple Silicon, llmfit treats system RAM as a shared CPU/GPU pool.
If a value is wrong, fix detection or use the documented hardware overrides before choosing a model. A recommendation calculated from an incorrect VRAM figure is not useful, however polished its score looks.
Step 3: open the model-fit interface
Launch the interactive terminal interface:
llmfitThe system summary appears at the top. Models are ranked below it with their score, estimated tokens per second, memory percentage, quantisation, context and run mode.
Current llmfit documentation also describes an automatically started web dashboard. If you only want the terminal interface, launch llmfit --no-dashboard. If you leave the dashboard enabled, check the address printed at startup and your firewall rules; do not deliberately bind it to 0.0.0.0 unless you intend to make it reachable from other devices.
These are the most useful controls for a first pass:
- Press
fto restrict the list to models marked runnable, good or perfect for the detected hardware. - Press
Uto choose a use case such as coding rather than accepting one generic ranking. - Press
Pto filter by provider, andRto filter by runtime/backend such as llama.cpp, MLX or vLLM. - Press
Lto inspect licence categories before using a model for paid or business work. - Press
/to search for a model family. - Press
Enterto open the selected model’s details. - Press
mto mark one model, thencto compare it with another.
Start with the fit result and memory percentage. A slightly lower-quality model with comfortable headroom is usually a better daily tool than a larger model that leaves no room for its context cache, the operating system or another application.
Step 4: inspect the recommendation instead of trusting the score
From a normal terminal, you can ask llmfit for one model’s assumptions:
llmfit info "<model name shown by llmfit>"Check four things:
- Quantisation: a Q4 file and an FP16 file from the same family have very different memory requirements and may not behave identically.
- Context: allocating the advertised maximum can consume much more memory than an 8K or 16K starting context.
- Run mode: full GPU, CPU offload and CPU-only operation can all "fit" while producing very different response times.
- Estimate source: prefer a measurement from the same hardware and runtime over a generic formula, then reproduce it yourself.
You can also obtain machine-readable recommendations for a script:
llmfit recommend --jsonThat is useful for inventory reporting, but do not automatically download or deploy a model based on one composite score. Include a human check for licence, storage, data sensitivity and intended workload.
Step 5: connect Ollama or llama.cpp
Use Ollama for the easiest path
Install and start Ollama separately. llmfit looks for Ollama’s local API at http://localhost:11434. Relaunch llmfit or press r to refresh providers.
Select a supported model and press d. If more than one provider can fetch it, choose Ollama in the provider window. llmfit then asks Ollama to pull the model and shows the progress.
If you are considering Qwen3.8 specifically, read the site’s Qwen3.8-27B local guide before committing storage and memory. A 27B model is not a sensible first download for every PC merely because a quantised variant appears in the catalogue.
Use llama.cpp for direct GGUF control
Place llama-cli or llama-server in your PATH. llmfit can download compatible GGUF files into its model cache, which defaults to ~/.cache/llmfit/models. Press D in the TUI to inspect or change that directory.
If llama.cpp is elsewhere, set LLAMA_CPP_PATH to the folder containing its binaries before launching llmfit.
For readers who would rather manage models from a graphical interface, the site’s Unsloth Desktop setup guide covers a different local-model workflow.
Step 6: replace the estimate with a real benchmark
Download a modest model first and serve it through the chosen runtime. Then return to llmfit, press r and select the installed model.
Press b. When llmfit detects both the model and its running provider, it offers to benchmark the model before opening the community leaderboard. Confirm the test and let it run three inference passes. The tool records tokens per second and time to first token.
Benchmark results are stored locally first. Sharing them to GitHub is optional and requires a separate choice. Leave sharing off until you have reviewed what the project submits and you are comfortable linking the result to your GitHub account.
A synthetic prompt is still not the whole answer. Test the model on one real task: a code edit, document extraction, private note summary or another representative workload. Check accuracy and latency together.
How to update llmfit
Use the same package manager that installed it:
# Windows / Scoop
scoop update llmfit
# macOS or Linux / Homebrew
brew update
brew upgrade llmfit
# uv
uv tool install -U llmfitRun llmfit --version afterwards. Update before assessing a model released after your installed catalogue, and read the release notes when provider detection changes.
Common llmfit problems
llmfit is not recognised
Close and reopen the terminal so it receives the updated PATH. For an isolated uv install, confirm the tool directory is in your user path. Run the package manager’s list command to verify that installation completed.
The GPU or VRAM is wrong
Run llmfit doctor. Confirm that the operating-system utility used for detection is installed. Windows currently documents automatic NVIDIA detection through nvidia-smi; native AMD and Intel GPU reporting is not equivalent. Use llmfit’s --memory, --ram or --cpu-cores overrides only after verifying the real hardware values.
A model marked "fit" runs out of memory
Reduce the context, close other GPU-heavy applications and choose a smaller quantisation or model. Runtime buffers, context cache and concurrent requests consume memory beyond the weight file. Treat "fit" as a shortlist, not a warranty.
Ollama is not detected
Confirm that Ollama is running and that http://localhost:11434 is reachable locally. If it runs on another host, set the documented OLLAMA_HOST value before starting llmfit. Do not expose an unauthenticated Ollama endpoint directly to the public internet.
llama.cpp is not detected
Make sure llama-cli or llama-server is in PATH. Otherwise set LLAMA_CPP_PATH to its binary directory, restart llmfit and press r.
A newly released model is missing
Update llmfit and check its current release notes. The catalogue may legitimately lag a model release, or the model may not yet have a supported local format. Do not substitute an unofficial file merely to make the name appear.
llmfit FAQ
Does llmfit run local models?
Not by itself. It recommends models and integrates with runtimes such as Ollama, llama.cpp, MLX, Docker Model Runner and LM Studio. A runtime performs the actual inference.
Are llmfit’s speed figures exact?
No. Some figures are estimates based on hardware and model characteristics; others may come from measured community results. Benchmark the exact model, quantisation and runtime on your machine.
Is llmfit free?
Yes. The project is released under the MIT licence. Individual model weights have their own licences, which may impose separate restrictions.
Does llmfit work on Windows with an AMD GPU?
RAM and CPU detection work, but the current platform guide documents automatic Windows GPU detection through NVIDIA’s nvidia-smi. Verify AMD memory manually and use a documented override for planning. That override changes the recommendation calculation; it does not add runtime support.
Does llmfit upload my hardware information?
The project says it does not contact external network services unless you explicitly use a feature that requires it, such as model downloads, runtime-provider queries or the community leaderboard. Benchmark sharing is optional. The terminal interface can also start a bundled local web dashboard; that is a listener rather than an upload, but you should check its bind address or disable it with --no-dashboard. Review the current documentation and network behaviour before using llmfit on a sensitive system.
Should you install llmfit?
llmfit is worth installing if you regularly ask whether a new local model will fit your PC, Mac or Linux workstation. It is much better than guessing from parameter count alone because it brings quantisation, context, run mode and the available hardware into one comparison.
The safest workflow is simple: verify detection, filter to a realistic use case, inspect the recommendation, download a modest model through a trusted runtime and benchmark it locally. That turns llmfit from a model-ranking screen into a practical purchasing and deployment aid.