Check the confidence label before trusting the speed column
A tokens-per-second figure in llmfit can come from a benchmark on your machine or from a formula. The scoring guide identifies those cases explicitly. That distinction should decide how you use the recommendation: an estimate is enough to investigate a model, while a hardware purchase needs evidence from the workload you intend to run.
llmfit reads your hardware and ranks models for memory fit, speed, quality and context. You can browse the terminal interface or request JSON for a script. The README also documents a web interface and API. I'd start with the terminal on the intended machine, where you can compare the detected hardware with what you know is installed before accepting the ranking.
What happened when we ran it
Our llmfit run passed 2,042 tests with no failures in 210 seconds. Installation succeeded in 14 seconds and recorded 237 dependencies, followed by a successful 140-second build. We ran commit e897fa5 on October 13, 2026, in a sandbox configured with 3 CPUs, 12g RAM and image lab-rust:1. The image label does not establish the actual Rust toolchain or operating system.
The checkout contained 678 files, roughly 77,307 lines of source and 29.7 MB of repository content. The scan found 7 CI workflow files, a Dockerfile and a tests directory. These are useful repository facts alongside the completed code checks. They do not establish that a particular GPU was detected correctly or that a suggested model will load at your desired context length.
The supplied run contains no inference throughput measurement. It cannot validate the speed column, compare accelerator backends or tell you which model answers your questions well. The build and test results justify trying the tool's selection workflow; any claim about model performance needs a different experiment.
The estimate exposes its assumptions
The confidence label separates local measurements from community results, calibrated formulas and plain estimates. An unsupported path receives its own label. Check that label along with the model's reported speed instead of comparing numbers whose evidence differs.
For recognized GPUs, the documented speed model uses memory bandwidth and model size with an efficiency factor. Unrecognized GPUs can fall back to backend constants. The guide also describes separate handling for prompt processing, so an estimated generation rate does not automatically establish how long you'll wait for the first token. These are documented mechanisms, not inference behaviors validated by our sandbox.
Run llmfit info with the selected model name to inspect its estimate basis and verification commands. This is a suggested next step from the README, not an interaction reproduced in our lab run. If the hardware identity or memory assumption looks wrong, resolve that before letting the ranking influence a download or purchase.
Quality scores still need your own task examples
The scoring guide derives quality from factors including parameter count, model-family reputation, quantization and task alignment. It uses a curated family benchmark table and falls back to name-based heuristics for families without an entry. That can help sort a large catalog, but it does not establish that a model will handle your private documents or coding conventions.
Choose a candidate that fits, then try examples drawn from the actual job. A coding assistant and a document extractor can disagree on the useful winner even on the same machine. llmfit helps with resource selection; you still own the acceptance criteria for the answers. Keep the model format and context requirement attached to the candidate rather than comparing family names alone.
Catalog updates require an updated executable
The model database is embedded at compile time, according to the scoring guide. End users receive updated catalog content by upgrading llmfit. If a model you expect is missing, check the tool version and catalog before assuming the machine cannot run it. A missing catalog entry and an incompatible model are different findings.
Platform support also has explicit limits. The platform guide says Android GPU autodetection is unsupported. Its manual memory workaround affects recommendations and does not add mobile GPU detection. Windows users should separately verify release signatures: the README warns that a release can publish an unsigned binary when the signing job fails or is skipped.
Start a provider when you are ready to measure
Ollama exposes an HTTP API for running and managing models. Its README also documents direct model execution. llama.cpp provides command-line and server interfaces with multiple hardware backends. If you've already chosen the model, going directly to one of these runtimes may answer your question sooner. llmfit becomes useful when you still need to compare candidates or inspect why one is recommended.
The source-setup rating is 4/5 for completed install, build and test stages in one sandbox, with desktop and GPU setup unassessed. Documentation gets 4/5 for explaining assumptions and platform limits, without a reproduced benchmarking workflow. Community and hardware maturity remain unassessed; repository activity and a passing suite cannot fill those gaps.
For a final decision, follow the README's bench workflow against a running provider and downloaded model. Keep the resulting measurement tied to that hardware, runtime and workload. If you're choosing what to download tonight, a fit estimate may be enough to proceed. Before spending money on a machine, require a measured result.

