Inferno: Local AI, Offline

Inferno is a private AI assistant that runs entirely on your phone. Chat with a local language model, ask questions about your photos and generate images, all on the device's own processor. Nothing leaves the phone: no account, no cloud, no tracking.

OFFLINE AND PRIVATE
• The only time Inferno uses the network is when you choose to download a model. Everything else works in airplane mode.
• Chats, photos and generated images stay in the app's private storage, excluded from backups and device transfers.
• No sign-up, no telemetry, no analytics, no ads. There is nothing to opt out of.
• Open source under the GPL-3.0. The code is on GitHub.

AI CHAT
• Streaming answers with real Markdown: headings, lists, tables and syntax-highlighted code with one-tap copy.
• Thinking mode: watch the model reason before it answers, then expand or collapse its thoughts.
• Edit and regenerate any turn.
• Long conversations keep going. When a chat outgrows the model's memory, choose rolling history, automatic summaries or a clean stop.
• Answers keep generating in the background while you use other apps.

VISION
• Attach a photo from the gallery, the camera or the share sheet and ask about it: identify objects, read text, describe a scene, turn a photo of the fridge into a recipe.
• Three photo-quality levels trade speed for detail.

IMAGE GENERATION
• Two on-device Stable Diffusion models: DreamShaper 8 LCM for quality, SDXS-512 for speed.
• Watch each picture form step by step, then save it, share it or attach it to a chat.
• A private gallery of everything you have made.

NINE CURATED MODELS, ALL WITH VISION
Qwen3.5 (0.8B, 2B, 4B), Gemma 4 E2B, MiniCPM-V 4.6, LFM2.5-VL (450M, 1.6B, 3B) and Qwen3-VL 2B, from 0.5 GB to 3.9 GB. Each card shows size, speed tier and licence before you download. You can also import your own GGUF files.

CHOOSE YOUR LEVEL OF CONTROL
• Normal: just chat. Sensible defaults, nothing technical.
• Power user: context length and policy, system prompt, sampling, thinking, performance presets.
• Developer: threads, core pinning, KV cache, flash attention, a built-in benchmark and per-turn timings.

BUILT FOR PHONES
Inferno runs llama.cpp and stable-diffusion.cpp with Arm KleidiAI kernels, threads pinned to the big cores and thermal-aware throttling. On Qualcomm phones the language model runs on the Adreno GPU. A Material 3 Expressive, monochrome interface keeps the focus on your words.

REQUIREMENTS
• Android 13 or newer, 64-bit ARM.
• 6 GB of RAM minimum, 8 GB recommended.
• A CPU with the dotprod and fp16 extensions (most phones from 2020 onward). Inferno checks at start-up and tells you if the phone is not supported.
• Models are downloaded from Hugging Face on demand and are not included in the app. Each model's licence is shown before download.

Inferno is made by Eyed.

Source code and issue tracker: github.com/cesp99/Inferno