LemoLM

Open LemoLM and choose New Chat to begin with the active model shown on the home screen. Enter a prompt and LemoLM will generate a response using the selected local engine. Open Models at any time to compare or change models.

Open the Models screen to view supported local model options. Model cards show information such as status, expected size, runtime, device suitability, licence, model type, and a model page link where available.

LemoLM is free to use with Mock Engine, Apple Foundation Models where available, Qwen3 1.7B, and SmolLM2 1.7B. The optional one-time Full Model Playground Unlock adds supported specialist, reasoning, coding, multilingual, and larger comparison models. Availability can change with licensing, runtime compatibility, and device capability.

LemoLM runs on compatible Apple silicon Macs. Mac hardware can provide more unified memory for larger models, but suitability still depends on model size and available resources. GPT-OSS 20B requires a Mac with at least 24 GB unified memory and is a large download intended for desktop-class local reasoning.

GPT-OSS 20B offers Low, Medium, and High reasoning effort. Higher effort may spend more time and output tokens before producing a final answer. If generation ends before an answer appears, reduce the conversation context, increase the output allowance, or choose a lower reasoning effort.

LFM2 English-Japanese Translation is a compact specialist model for focused English-to-Japanese and Japanese-to-English translation. It is intended for short translation tasks rather than general conversation or factual question answering.

Choose Precise, Balanced, or Creative for quick setup, or use Custom to adjust Temperature, Top-p, Top-k, and Seed. Generation and context settings persist between sessions and can be reset to their defaults.

Performance Snapshot records recent generation timing together with available device, operating-system, thermal, power, battery, Low Power Mode, and memory context. Conversation sharing can export the complete snapshot and conversation, the snapshot only, or the conversation only as Markdown.

Open Load Diagnostics when a downloadable model does not load as expected. It shows device memory, runtime mode, model-file status, and recent generation information that can help identify compatibility or memory-pressure problems.

LemoLM is local-first. Conversations are stored on your device, and supported local engines process prompts and responses on-device. Model downloads require internet access and may connect to third-party model hosts such as Hugging Face.

Compact on-device language models can hallucinate, misunderstand prompts, omit context, or provide inaccurate information. Always verify important information, especially for health, legal, financial, safety, or current-event topics.

Performance depends on your device, available memory, operating-system version, thermal and power state, and selected model. Larger models may download or load slowly, use more energy, heat the device, or fail on lower-memory hardware. Compact models are generally better suited to older or lower-memory devices.