Free vs Pro

Free vs Pro

llama-android follows an Open Core model. The Free tier is a complete, production-ready on-device LLM runtime, MIT-licensed and public. Pro adds performance and advanced-inference features for teams that need them.

Feature	Free	Pro
Chat completion (complete)	✅	✅
Embeddings (embed)	✅	✅
System prompt + auto chat template	✅	✅
GGUF, any quantization	✅	✅
CPU / NEON	✅	✅
Vulkan GPU acceleration	✗	✅
Streaming tokens (Kotlin Flow)	✗	✅
JSON / grammar-constrained output	✗	✅
Vision (multimodal, image input via libmtmd)	✗	✅
Context save / restore (persist a conversation)	✗	✅
Multiple concurrent sessions (per-session KV cache)	✗	✅
LoRA adapters (hot-swap fine-tunes)	✗	✅
Function calling (tool use)	✗	✅
ABI	arm64-v8a	arm64-v8a + x86_64
Channel	Maven Central · JitPack · GitHub Release	Gumroad

The Free API

The Free tier exposes exactly five methods — everything you need for on-device chat and embeddings:

Llama.loadModel(path, config)   // load a GGUF model
Llama.complete(model, prompt)   // chat completion → LlamaResult
Llama.embed(model, text)        // text → FloatArray
Llama.getSystemInfo()           // CPU / backend info
Llama.releaseModel(model)       // free native memory

What Pro adds

• Vulkan GPU — offload layers to the GPU for multi-× faster generation on capable phones (gpuLayers > 0).
• Streaming — receive tokens as they're produced via a Flow<String>, for a live-typing UX.
• JSON / grammar — constrain the model to valid JSON or a GBNF grammar so the output always parses.
• Vision — feed images to multimodal models (LLaVA-style) via libmtmd.
• Context save / restore — snapshot and reload a conversation's KV cache to resume chats across app restarts.
• Multi-session — run several independent conversations concurrently, each with its own KV cache.
• LoRA adapters — load and hot-swap LoRA fine-tunes at runtime.
• Function calling — structured tool-use for agent-style apps.

→ Get llama.cpp Pro

Pro is distributed only via jokobee.com — never on Maven Central or JitPack.