ikppramesh/irx-mini-GGUF

IRx-mini (GGUF)

The smallest member of the IRx family: a private, offline AI chat assistant made to run on phones with just 4GB of RAM. Everything runs on your device — no internet connection, no account, and nothing you type ever leaves your phone. This is the build for phones, tablets and llama.cpp-based apps (PocketPal, LM Studio, llama.rn, Ollama).

Download: irx-mini-Q8_0.gguf (~1.1GB) — one file for iPhone, iPad, Android and Mac. Uses roughly 1.3GB of memory while running, so it fits phones with 4GB of RAM.

The IRx family

All four run fully offline — nothing you type ever leaves your device.

Model	Download	Best for	Phones & tablets (GGUF)	Mac (MLX)
IRx-mini	~1.1GB	The smallest. Simple, quick answers on phones with just 4GB of RAM	irx-mini-GGUF	irx-mini
IRx-1	~1.2GB	Fast, quick everyday answers on almost any phone	irx-1-GGUF	irx-1
⭐ IRx-2 — Best Overall	~2.7GB	The best balance of answer quality and speed	irx-2-GGUF	irx-2
IRx-2 Pro	~5.3GB	The most detailed, long-form answers and long conversations	irx-2-pro-GGUF	irx-2-pro

Not sure which to pick? Start with ⭐ IRx-2. Choose IRx-mini for a 4GB-RAM phone, IRx-1 for fast everyday answers on most phones, and IRx-2 Pro for the richest answers on a device with 12GB+ RAM.

See them side by side: the IRx board compares all four models, shows their measured speed live, and has an animated step-by-step setup guide for PocketPal on iPhone and Android.

What IRx-mini does best

IRx-mini is the smallest member of the family, made for phones with only 4GB of RAM. It gives short, simple answers, quick explanations and everyday help, fully offline, on phones that can't fit the larger models.

Who	Example things to ask
Farmers	"When should I water tomato plants?" · "Write a short message to my buyer."
Students	"What is photosynthesis? Explain simply." · "5 quick quiz questions on fractions."
Shop & small business owners	"Write a one-line offer for my shop." · "How do I calculate 15% profit?"
Software developers	"What does HTTP error 404 mean?"
Travellers	"Make a short packing list for 2 days."
Everyday life	"Convert 25°C to Fahrenheit." · "Suggest a quick breakfast."

Why a model for 4GB phones?

Many phones in use today, especially budget Android phones and older models, have only 4GB of RAM. The phone's own system and background apps already use a large part of that, so an AI model has to be small to load and run reliably — the larger IRx models are simply too big for these phones.

IRx-mini is built for exactly that:

• Small enough to fit. About 1.1GB to download and roughly 1.3GB of memory while running, leaving room for the phone's system and other apps.
• Works with no internet. Useful where connectivity is weak, slow or expensive — on farms, in villages, while travelling, or on a limited data plan. Nothing is sent to the cloud, so there are no data costs and no waiting on a network.
• Private. Everything stays on the phone.
• Light on storage and battery. A small model loads quickly and does less work per answer.

The trade-off is depth: IRx-mini gives shorter, simpler answers and makes more mistakes than the larger models. If your phone has 8GB+ RAM, ⭐ IRx-2 will give much better answers.

Measured on the same Mac CPU: IRx-mini generates about 89 tokens/s — the fastest of the family (IRx-1: about 51 in the same test).

Recommended settings (PocketPal and similar apps)

Setting	Value
Context size	8192
Max new tokens	1024
Temperature	0.7
Top-P / Min-P	0.95 / 0.05
Repeat penalty	1.1
System prompt	leave empty — IRx-mini's identity is built into the model

IRx-mini passed its loop tests even with the repeat penalty off (PocketPal's default), but 1.1 is still recommended. It is also embedded in the file as the default; apps that let you set your own sampling values (PocketPal does) need it set there too.

Updating from an older download: delete the old model in the app and download it again — apps don't replace a downloaded file on their own. If you set a custom chat template in the app, reset it so the one built into the file is used.

Built-in chat template

Apps built on llama.cpp often switch on a hidden "thinking" mode by default, which can use up the whole conversation before any answer appears ("The conversation ran out of room"). The template embedded in this file always keeps thinking off — the app's reasoning toggle has no effect. IRx-mini needs no system prompt: its name and creator are trained into the model itself, so nothing is added to your messages.

Usage (command line)

llama-cli -m irx-mini-Q8_0.gguf --jinja -c 8192 -p "How do I convert Celsius to Fahrenheit?"

Limitations

• Not a frontier-scale model — it won't match large hosted AI services on very hard multi-step reasoning, deep coding problems or breadth of world knowledge.
• Facts can be wrong. Like any small offline model it can state things confidently that aren't true — double-check anything important, such as prices, medicines and doses, laws, and local farming advice (seed varieties, chemical quantities, weather).
• Not a replacement for professionals — medical, legal and financial answers are general information only.
• Limited news knowledge. IRx-mini was trained on headlines up to 3 October 2026, but at this size it recalls news poorly and may even give the wrong cutoff date. For recent news, use ⭐ IRx-2 or check a news source.
• Don't enable native tool/function-calling in chat apps — never trained; plain chat is reliable.

Changelog

• 2026-10-03 — Rebuilt: no more looping on long answers. The first release could repeat a phrase endlessly on long, structured answers such as day-by-day trip itineraries in PocketPal, whose default repeat penalty is off. IRx-mini is now built on a stronger 1B base:
  • Long answers: tested with trip itineraries (Lakshadweep, Kerala, Rajasthan) with the repeat penalty off and on, plus repeated 8-turn chats: no loops.
  • Size and speed: the download is about 1.1GB (was 0.8GB), still a fit for 4GB-RAM phones, and it is faster: about 89 tokens/s on the Mac CPU test (was 60–74).
  • Answer quality: answers are noticeably more complete and better structured, including code.
  • Identity: IRx-mini's name and creator are now trained into the model, with no system prompt added to your messages.
  • News: the built-in news digest was removed from IRx-mini, because at this size the long prompt pulled answers off-topic. News recall is limited (see Limitations).
  • Updating: delete the old IRx-mini in your app and download it again.
• 2026-10-03 — First release, trained on news up to 3 October 2026 from 18 sources (see the news update notes on the other IRx cards). Built with the IRx pipeline's fixes: fine-tune merged into the full-precision base and quantized once with an importance matrix, thinking mode always off, identity built in. Passed the automated pre-publish check: repeated 8-turn conversations with the repeat penalty off and the app's thinking request on — no loops, no copied answers, no hidden reasoning, correct identity.

License

Apache 2.0. IRx-mini is a derivative fine-tuned model — full Apache 2.0 terms apply as with any work under this license.