The first obstacle when trying out local LLMs is not knowing which models your machine can actually handle. llmfit replaces that judgment call with a single terminal command. This Rust-based open-source tool automatically detects your RAM, CPU, and GPU (VRAM), then scores each model on whether it fits in memory, how fast it will likely run, and which quantization level you should choose. It is developed by Alex Jones (AlexsJones) in the UK and has been updated at close to a weekly pace since its release in February 2026. It is MIT licensed and free to use.
Key Features
- Automatic hardware detection: On startup it identifies installed RAM, CPU, GPU/VRAM, and available backends. Running
llmfit doctorproduces a detection report, which also helps troubleshoot cases where hardware is misidentified - Four-axis fit scoring: Models are scored on memory fit, estimated speed, quality, and context length, then ranked. You see not just whether a model runs, but whether the combination is practical
- Quantization suggestions and speed estimates: It proposes quantization levels suited to your hardware (GGUF Q4/Q5/Q8 and so on) and gives a tokens-per-second (tok/s) estimate. Estimates are based on a memory-bandwidth model informed by community measurements
- Support for major local runtimes: It detects Ollama, llama.cpp, MLX, LM Studio, Docker Model Runner, and RamaLama, including models you have already installed
- Correct handling of MoE and multi-GPU: Mixture-of-Experts models are easy to misjudge because total parameters differ from what is actually active, and llmfit accounts for this in its scoring. Multi-GPU setups are supported as well
- Both TUI and CLI: Launching with no arguments opens an interactive TUI for filtering models, downloading, and running benchmarks. CLI modes such as
llmfit fitandllmfit recommend --jsonare also available for scripting - Benchmarks and leaderboard:
llmfit benchmeasures real performance on your machine and lets you contribute the result to a community leaderboard, so estimate accuracy improves through user measurements
Pricing
| Plan | Price | Highlights |
|---|---|---|
| Open source | $0 (MIT license) | All features free to use. No paid tier offered |
Pricing is accurate as of August 2026. Check the official site and the GitHub repository for the latest information.
Several installation routes are provided. On macOS/Linux use brew install AlexsJones/llmfit/llmfit; on Windows use scoop install llmfit. MacPorts, uv tool install -U llmfit (via Python), a Docker image (ghcr.io/alexsjones/llmfit), an install script, and building from source with cargo build --release are also available.
Pros & Cons
✅ Pros
- Answers the first question — “which models run on this PC?” — in a matter of seconds
- Lets you decide before downloading, avoiding wasted downloads of tens of gigabytes
- A single Rust binary, with no runtime or dependencies to prepare
- Not tied to one runtime; it looks across Ollama, llama.cpp, MLX, and others
- Actively maintained, with quick support for new model families and hardware
⚠️ Cons
- Terminal-oriented, which raises the barrier for GUI-only users
- Speed figures are estimates. Actual numbers vary with quantization format, context length, and concurrency, so running a benchmark is needed for accuracy
- As an individually maintained open-source project, it has no vendor support desk or long-term guarantees
- It does not run inference itself; it focuses on diagnosis and selection, leaving execution to Ollama and similar tools
Comparison with Similar Services
| Criteria | llmfit | Ollama | LM Studio | Hugging Face Model Memory Calculator |
|---|---|---|---|---|
| Main role | Model selection diagnosis | Local inference runtime | GUI local inference environment | Rough memory requirement estimate |
| Interface | TUI / CLI | CLI / API | Desktop GUI | Web form |
| Automatic hardware detection | Yes (RAM/CPU/GPU) | Limited | Yes (shows rough compatibility) | No (manual input) |
| Speed estimation | Yes (tok/s estimate + real benchmarks) | No | No | No |
| Quantization suggestions | Yes | No | Partial | No |
| Price | Free (MIT) | Free | Free (personal use) | Free |
llmfit is not a replacement for a runtime; it is easier to place if you think of it as a diagnostic layer that sits in front of one. The intended pattern is to narrow down candidates with llmfit before running them in Ollama.
Who Is It For
- People starting with local LLMs who have no idea what their current PC can run
- Existing local LLM users tired of downloading model after model to find out
- Anyone checking which class of model comes within reach before buying a new machine or GPU
- Developers who want a consistent basis for model selection across several in-house machines
- Engineers who want to automate per-environment model selection from CI or scripts using JSON output
Summary
llmfit moves the first hurdle of local LLM adoption — picking a model that fits — from guesswork to numeric evaluation. Automatic hardware detection, four-axis fit scoring, quantization suggestions, and tok/s estimates together give you the information you need before downloading anything. Since it does not perform inference, it assumes pairing with Ollama or llama.cpp, but it is free and ships as a single binary, so the cost of trying it is low. It is worth a run before you dive into local LLMs.