AGENTCONN
L

// Agent profile

llmfit

DevOps Free

About llmfit

llmfit is a Rust-built command-line tool that solves the 'will this model even run on my machine?' problem. It detects your system's RAM, VRAM, GPU, and CPU capabilities, then scores hundreds of models across memory fit, estimated speed, quality, and context window size. An interactive TUI lets you browse rankings, simulate configurations, and download models directly. Supports Ollama, llama.cpp, MLX, Docker Model Runner, and LM Studio as local runtimes. The --json flag enables agent and CI integration. Community-contributed model profiles ship with each release. MIT licensed, 33K+ stars.

Key Features

  • Auto-detects RAM, VRAM, GPU, and CPU — no manual spec entry
  • Multi-dimensional scoring: memory fit, estimated speed, quality, context window
  • Interactive TUI with model ranking, planning simulation, and direct download
  • Benchmarking mode — measures actual tokens/sec throughput on your hardware
  • Multi-GPU and MoE architecture support (Mixtral, DeepSeek-V3)
  • Supports Ollama, llama.cpp, MLX, Docker Model Runner, LM Studio
  • JSON output for agent consumption and CI/CD pipeline integration
  • Community-maintained model catalog — new profiles ship with each release
  • Cross-platform: Homebrew, Scoop, pip/uv, Docker, Cargo from source

Overview

llmfit answers the most common question in local AI: “which model should I actually run?” Instead of guessing based on parameter counts and VRAM tables, llmfit scans your hardware — GPU model, available VRAM, system RAM, CPU — and produces a ranked list of models that will actually fit and perform. The scoring is multi-dimensional: memory fit determines if it runs at all, estimated speed shows throughput, quality reflects benchmark performance, and context window size factors in for agentic workloads that need long contexts.

The tool is particularly relevant for agent operators choosing local models for coding agents, RAG pipelines, or tool-use loops where token throughput and context window matter as much as raw quality scores.

Key Capabilities

Hardware detection and scoring: llmfit identifies your system’s capabilities automatically — no manual entry of GPU model or RAM size. It then evaluates each model in its catalog against your hardware profile, producing a composite score. The speed estimation uses a memory-bandwidth model grounded in community benchmarks, with full transparency: llmfit info <model> shows every assumption so you can verify.

Interactive TUI: The terminal interface lets you browse ranked models, simulate different quantization levels, compare multiple models side-by-side, and download directly to your local runtime. For automation, llmfit recommend --json outputs structured data for agent consumption.

Benchmarking: Beyond estimates, llmfit bench measures actual tokens-per-second throughput on your hardware. This matters for agent workloads where the difference between 15 t/s and 40 t/s determines whether tool-use loops feel responsive or painful.

Use Cases

Agent operators selecting local models for coding agents, MCP servers, or autonomous loops. DevOps teams standardizing on models across heterogeneous hardware (different GPU generations, M-series Macs, CPU-only servers). CI/CD pipelines that need to validate model compatibility before deployment. Developers evaluating whether to upgrade hardware or switch quantization levels to fit a target model.

Considerations

llmfit recommends models — it doesn’t run them. You still need a local runtime (Ollama, llama.cpp, MLX, etc.) installed separately. The model catalog depends on community contributions, so newly released models may take a few days to appear. Speed estimates are approximations based on memory bandwidth modeling; the bench command gives ground truth but takes time to run. Privacy-conscious: no data leaves your machine unless you explicitly use download or leaderboard features.

Who It’s For

Anyone running local LLMs who wants data-driven model selection instead of guesswork. Agent developers who need to match models to hardware constraints for production deployments. Teams managing fleets of machines with different GPU/RAM configurations who need consistent model recommendations across the fleet.

Similar Agents