LLMario 0.2.1: a more accurate memory check

A more accurate memory check, so more models fit. LLMario is a free, open-source app for chatting with open models (Qwen, Gemma, Llama, gpt-oss, Mistral and more) on your own computer. Nothing you type leaves your machine.

Downloads

	File
macOS 13+ (Apple Silicon and Intel)	LLMario-0.2.1-macos-universal.dmg	Open it and drag LLMario into Applications. Signed and notarized by Apple.
Windows 10/11 x64: app	LLMario-0.2.1-windows-x64-setup.exe	Installs for your user account; no admin rights needed.
Windows 10/11 x64: command line	llmario-0.2.1-windows-x64.zip	llmario.exe (CLI and local OpenAI-compatible API). Unzip anywhere.

Each file has a .sha256 next to it. To check a download:

• macOS / Linux: shasum -a 256 -c <file>.sha256
• Windows (PowerShell): Get-FileHash <file> and compare with the .sha256 file

Windows preview: SmartScreen warning

The Windows files are not code-signed yet. The first time you run the installer, Windows may show "Windows protected your PC". Click More info → Run anyway. Code signing is planned.

Before you start: install an engine

LLMario runs models with the open-source engine llama.cpp (and MLX-LM on Apple Silicon Macs). The app does not include an engine yet, so install one:

• Windows: winget install ggml.llamacpp in a terminal, then restart LLMario
• macOS: brew install llama.cpp, and optionally pip3 install mlx-lm on Apple Silicon (usually fastest)

The app's Models → Engines tab shows what it found.

What's new in 0.2.1

• More models fit in limited memory. The memory check now follows each model's real layout.
  • Only the layers whose memory grows with the conversation are counted that way. Sliding-window layers stop at their window, and recurrent layers use a fixed amount.
  • Example: on a 16 GB Mac, Gemma 4 12B (GGUF) now plans at 10.2 GiB. 0.2.0 planned 14.6 GiB and refused to load it.
  • Models you added before 0.2.1: run llmario model refresh once to give them their layout.
  • To go back to the old, more cautious estimate, set [runtime] kv_accounting = "conservative" in the config.
• Less memory for llama.cpp's saved conversation states. llama.cpp now keeps 2 of them instead of up to 32, with the same reuse of earlier prompts. For Gemma 4 12B in a chat, that is 0.66 GiB instead of 2.54 GiB.
• Memory readings on macOS include the model itself. llama.cpp maps model files into memory, and those pages were missing from the numbers in the app, llmario doctor and the local API's metrics. For gpt-oss-20b, 0.5 GiB was shown instead of 11.5 GiB.
• Models on external drives: the hidden ._ files macOS creates on exFAT drives are no longer recorded as part of a model, so llmario model verify no longer reports false problems. llmario model refresh removes such records from models added earlier, without deleting any files.
• LLMario Beta can run alongside. LLMario warns when the beta has models loaded, since both share your computer's memory.
• Benchmarks: llmario bench refuses to run on a busy computer (--allow-busy overrides), warms up every prompt length and reports effective memory bandwidth.

Status

• Memory estimates: checked on Apple Silicon (M4 Max). The estimates stayed 8–16% above the measured peaks for Qwen3.5 9B, Gemma 4 12B and gpt-oss-20b. The 16 GB figures above are computed, not yet measured on a 16 GB Mac.
• Windows (preview): built and tested automatically on Windows (install and launch of the app, a real chat through llama.cpp, clean engine shutdown). Please report issues.
• macOS: the Intel part of the universal build is untested on real Intel hardware.

Website: https://llmario.com · Source: https://github.com/mannysinghx/LLMario