LLMForge
Orchestrate your LLM pipeline. Locally
Last verified:
What is LLMForge?
LLMForge is a native macOS application that provides the complete LLM pipeline in one app, allowing users to fine-tune and ship LLMs directly from their Mac without using the terminal or cloud services. It handles the full workflow from model download to on-device deployment in a single native window, making it ideal for developers who want to avoid stitching together CLI tools.
Key features include browsing and downloading models from HuggingFace with visibility into architecture, size, and RAM requirements before downloading; curating training data by importing CSV/JSONL files, manually labeling, or having a local model generate pairs; fine-tuning natively on Apple Silicon using MLX without CUDA or cloud GPUs; quantizing and exporting to GGUF or CoreML formats; testing models side-by-side with simultaneous responses; and serving locally via an OpenAI-compatible API at localhost for instant app testing.
LLMForge is specifically designed for macOS developers with Apple Silicon (M1+) chips who want to work with LLMs privately and offline. It supports models like Phi-3 Mini, Llama 3.2 3B, Qwen 2.5 1.5B, and Gemma 3 1B. The app is RAM-aware with clear warnings before operations, exports to Ollama and LM Studio with one-click, and includes a built-in model library for managing local models with tagging and version tracking.
LLMForge pricing
Pricing model: Freemium
Free tier available for models under 200 MB with no account required. No paid plan details explicitly stated on the website. The app itself is free to download and use for qualifying model sizes. Requires Apple Silicon (M1+), macOS 26+, and 8 GB RAM.
LLMForge pros
- 100% local and private - data never leaves your machine
- No terminal required - complete GUI workflow
- No cloud dependency - fully offline operation
- Native MLX Apple Silicon training - no CUDA needed
- GGUF App-ready export for Xcode integration
- Built-in HuggingFace model browser with pre-download specs
- RAM-aware with clear OOM warnings before operations
- One-click export to Ollama and LM Studio
- OpenAI-compatible local API server at localhost:8080
- Side-by-side model testing with simultaneous responses
- Real-time loss curve visualization during fine-tuning
- Curate training data without writing scripts
- Auto-outputs clean Alpaca or ChatML format
- Quantization level picker balancing size vs quality
- Built-in model library with tagging and version tracking
- No account required to use
- Free under 200 MB model size
- Import CSV/JSONL training data directly
- Local model can generate training pairs you accept/reject
- Save great outputs back to dataset for next training run
LLMForge cons
- Requires Apple Silicon (M1+) only - no Intel Mac support
- Requires macOS 26+ - very new macOS version needed
- Requires 8 GB RAM minimum - limited for small devices
- Only 200 MB free tier - paid plan needed for larger models
- No cloud GPU option - entirely local hardware dependent
- No Windows or Linux support - macOS only
- Limited to MLX framework - no PyTorch direct training
- No distributed training across multiple machines
- GGUF export only - no other quantization formats besides CoreML
- Beta stage - may have unstable features
Frequently asked questions about LLMForge
What is LLMForge?
LLMForge is a native macOS application that provides the complete LLM pipeline in one app. It allows users to fine-tune and ship LLMs directly from their Mac without using the terminal or cloud services, handling the full workflow from model download to on-device deployment in a single native window.
Does LLMForge require cloud or internet access?
No, LLMForge is fully offline and private by default. Your data, models, and training never leave your machine. While you can browse HuggingFace to download models, the actual training and inference happen 100% locally on your Mac.
What Macs does LLMForge support?
LLMForge requires Apple Silicon (M1+) chips only. It does not support Intel Macs. The app also requires macOS 26+ and a minimum of 8 GB RAM.
How do I fine-tune a model in LLMForge?
To fine-tune: 1) Browse and download a model from HuggingFace within the app, 2) Curate training data by importing CSV/JSONL or generating pairs with a local model, 3) Configure LoRA rank, learning rate, and epochs, 4) Start fine-tuning which runs natively on MLX on your Apple Silicon, watching the loss curve descend in real time.
What quantization formats does LLMForge export?
LLMForge exports to GGUF and CoreML formats. For GGUF, you can choose quantization levels including Q8_0 (best quality, 7.2 GB), Q4_K_M (recommended, 3.8 GB), and Q2_K (smallest, 2.1 GB).
Can I test my fine-tuned model against the base model?
Yes, LLMForge has a side-by-side testing feature where you can use the same prompt with both the base model and fine-tuned model simultaneously, comparing quality and speed. You can see tok/sec metrics for each (e.g., 12.4 tok/sec vs 11.8 tok/sec).
How do I use my fine-tuned model in my application?
LLMForge can spin up an OpenAI-compatible API server from any fine-tuned model at localhost:8080/v1/chat/completions. Point your app at this localhost endpoint to test with real requests and iterate in seconds without any deployment needed.
Can I export my model to Ollama or LM Studio?
Yes, LLMForge offers one-click export to run your model in Ollama or LM Studio. There is no manual GGUF wrangling required - just click and export.
What models can I download through LLMForge?
You can search and download the entire HuggingFace model hub. Examples shown include Phi-3 Mini (3.8B params, 2.2 GB), Llama 3.2 3B (2.0 GB), Qwen 2.5 1.5B (1.0 GB), and Gemma 3 1B (0.7 GB). The app shows architecture, size, and RAM requirements before download.
Is LLMForge free to use?
LLMForge is free for models under 200 MB with no account required. The website states 'Free Under 200 MB No account required' at the bottom. No paid plan pricing details are explicitly provided on the website.