IRx-1

A small, fast AI chat assistant that runs entirely on your device — no internet connection, no account, and nothing you type ever leaves your phone, tablet or laptop. IRx-1 is the lightest member of the IRx family: quick everyday answers, even on modest phones.

The IRx family

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

Model	Download	Best for	Phones & tablets (GGUF)	Mac (MLX)
IRx-1	~1.2GB	The fastest. 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-1 for speed or an older / smaller phone, and IRx-2 Pro for the richest answers on a device with 12GB+ RAM.

What IRx-1 does best

IRx-1 is the smallest and fastest member of the family. It is built for quick, everyday help: short answers, simple explanations, quick drafts and checklists, instantly and offline — including places with no or poor internet.

Who	Example things to ask
Software developers	"What does HTTP error 404 mean?" · "Write a Python one-liner to reverse a list."
Farmers	"Give me a simple way to track spraying dates for my crops." · "Draft a short message to my buyer about tomorrow's delivery."
Students	"Explain photosynthesis in two sentences." · "Give me 5 quiz questions on fractions."
Teachers	"Suggest 3 fun warm-up activities for a class of 10-year-olds."
Shop & small business owners	"Write a one-line WhatsApp message for my weekend sale."
Travellers	"Make a packing checklist for a 3-day trip."
Everyday life	"Suggest a quick breakfast." · "Convert 25°C to Fahrenheit."

Limitations

• Not a frontier-scale model. At ~2B parameters, it won't match large hosted models on hard multi-step reasoning, deep technical/coding problems, or breadth of world knowledge — that gap is a function of scale, not something fine-tuning erases
• Not current-events aware on its own. The model's weights are fixed as of training — by itself it has no access to recent news, prices, or events. Fine-tuning more often on news doesn't fix this reliably (news is dense with exactly the kind of fast-changing, precise facts fine-tuning is worst at — see the next point); the fix that's actually reliable is retrieval, not retraining, which the reference repo provides as an opt-in add-on (see Strengths) — the base model file here has no built-in awareness of anything past its training data.
• Fine-tuning improves but doesn't guarantee facts. Tested directly: adding accurate, spot-checked training examples measurably improved general-knowledge accuracy (most capitals, chemical symbols, and historical facts now answer correctly), but individual facts can still be wrong — especially where the base model already had a strong competing belief from pretraining (e.g. it still answers "Sahara" for the largest desert despite a correct "Antarctica" training example, and gives an imprecise population figure for India despite an accurate one in training). This is a real, partial improvement, not a fix — treat factual claims as worth double-checking, same as any small model.
• Don't enable native tool/function-calling in chat apps. The base architecture supports a native <tool_call> function-calling format, but IRx-1's fine-tuning never trained on or reinforced it. If a host app (LM Studio, Bionic, or similar) exposes tools/functions to the model, it can misfire the wrong tool for a plain question, or spiral into an unterminated loop hallucinating a fake multi-step agentic session — especially in apps with a large always-on tool/skill system prompt, which can overwhelm a model this size regardless of the question asked. Plain chat (no tools exposed) is reliable.
• Occasional off answers at higher sampling temperatures — like any small model, it can occasionally produce a repetitive or degenerate response; regenerating usually resolves it
• Best suited to short-to-medium context — very long, multi-document context windows aren't its strong suit
• 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.

How it was built

Full technical detail, including the exact scripts and the story behind each data source, is in the GitHub repo.

Example queries

"Who created you?"

I was created by Ramesh Inampudi from Hyderabad, India. He built me, IRx-1, at iramesh.com.

"What questions should I ask before signing an apartment lease?"

1. What is the total rent, including security deposit and utilities?
2. What are the rent payment terms, such as due dates and early termination fees?
3. What specific appliances and furniture are included in the lease?
4. Are there any pet fees or restrictions on pets?
5. What are the maintenance responsibilities for both the landlord and tenant?
6. What are the rules regarding subletting or assigning the lease to a new tenant? ...

"What is the chemical symbol for gold?" / "Who painted the Mona Lisa?"

Au

Leonardo da Vinci

"Explain the difference between TCP and UDP in a couple of sentences."

TCP is a connection-oriented protocol that ensures reliable delivery by establishing a handshake, managing resources, and retransmitting lost packets, while UDP is a connectionless protocol that prioritizes speed over reliability, making it ideal for real-time applications like streaming and VoIP.

"Write a short, encouraging note to leave for a roommate who is stressed about exams."

Hey, I know your brain is working overtime! Just remember to breathe, and we'll tackle these problems one at a time. You've got this, and I'm right here cheering you on.

(Real, unedited outputs. Small models vary run to run — regenerate if a particular answer misses.)

Tool-use example: xGAIR

xGAIR is an MCP server that plugs AI coding assistants into any GitHub repo. Its chat CLI matched only exact command syntax; IRx-1 serves as an optional natural-language fallback, parsing free-form input into the correct structured tool call. Real, verified outputs, repo names not seen verbatim in training:

> hook up github.com/vercel/next.js
  → xgair_connect_repo { url: "github.com/vercel/next.js", repo: "next.js" }

> run discovery on stripe/stripe-node
  → xgair_discover_repo { repoId: "stripe/stripe-node" }

> check this snippet: DROP TABLE users;
  → xgair_validate { repoId: "", codeSnippet: "DROP TABLE users;" }

Known limitation: when a repo reference is embedded mid-sentence rather than at the start of the message, the extracted repoId sometimes comes back empty even though the tool choice itself is correct. Occasionally an off-topic message gets mapped to a tool call instead of {"tool": "unknown"}. Neither is catastrophic by design — the calling integration falls back to a current-repo context when repoId is empty, and a wrongly-triggered call is a harmless read, not a destructive action.

Current-events awareness: retrieval, not retraining

The weights above never change to add current-events knowledge — fine-tuning doesn't reliably teach new facts (see Limitations), and news is the worst case for that. The GitHub repo instead ships an opt-in RAG pipeline:

Covers Indian news and, since AI-model releases go stale even faster than general news, AI/tech news too — same reasoning, same mechanism. Two parallel paths from the same source data: the left one (SQLite + FTS5) only works on the Mac that built it. The right one is a free, static JSON snapshot (GitHub Actions + Pages, no server) reachable from anywhere:

https://ikppramesh.github.io/irx-1/news.json

This does not automatically make a mobile app "current." The URL above is a plain HTTP endpoint — something has to actually call it and search the result. That's real client-side code, not a property of the model file. Full working example (JS) in the GitHub repo.

Honest result, not oversold: narrow, specific questions ground correctly ("what did TechCrunch report about Meta's AI model?" answered accurately from the actual retrieved article). Broad, open-ended questions don't — "what are the latest AI models released?" still fell back to hallucinating stale, made-up model names from frozen training memory, even with relevant articles successfully retrieved and an explicit instruction to prefer them. A 2B model juggling several simultaneous instructions (identity + grounding + retrieved text) doesn't reliably prioritize all of them — the same pattern behind the tool-calling/agentic-overwhelm limitation below. Ask specific questions.

Usage (Mac, MLX)

pip install mlx-lm

from mlx_lm import load, generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load("ikppramesh/irx-1")
messages = [{"role": "user", "content": "How do I convert Celsius to Fahrenheit?"}]
prompt = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
print(generate(model, tokenizer, prompt=prompt, max_tokens=1024,
               sampler=make_sampler(temp=0.7, top_p=0.95)))

No system prompt is required: the built-in chat template supplies the IRx-1 one when none is given, and always keeps thinking mode off. Sample at a non-zero temperature (e.g. temp=0.7). On a phone or tablet, use the GGUF build.

Changelog

2026-09-30

• Fixed empty replies / "The conversation ran out of room" on phones. Apps built on llama.cpp (PocketPal and others) switch on the base architecture's thinking mode by default; the model then used its whole context on hidden reasoning before answering. The built-in chat template now always keeps thinking off (verified: 0 reasoning tokens with the app default on).
• Identity without a system prompt. No training example had ever taught the name — it came only from a system prompt most apps never send, so the GGUF answered the base model's name. Added identity training examples and a built-in default system prompt; verified answering as IRx-1 with no system prompt, and with a generic one.
• Retrained. Examples are now fitted to the training window by dropping the oldest turns (previously truncated mid-answer, which teaches stopping early), and loss is computed on answers only.
• Fixed the scheduled retrain: every run since 2026-09-11 had crashed with a GPU out-of-memory error at the 1024-token training window. Now 768 (9.7GB peak).
• Fixed repeating/looping replies in the GGUF build (a word or sentence repeated until the reply ran out, or an earlier answer pasted again). Cause: the GGUF was merged into the already-4-bit training base and quantized a second time; it's now merged into the full-precision base and quantized once with an importance matrix. An automated multi-turn loop check now runs before every publish.

2026-09-08

• Automated recurring retrain + publish pipeline added: every 5 hours, after the news-fetch cycle, the model retrains on a small rolling set of news-derived examples and republishes both this repo and the GGUF repo. Tested end-to-end successfully. Documented limitation carried forward unchanged: this does not reliably teach the model new facts (see below) — the actual current-events answer path remains retrieval, not this.
• Made news retrieval portable beyond the build Mac — free GitHub Actions + Pages publish a JSON snapshot any client (mobile included) can fetch and search with no server; verified live and serving real data
• Wired news retrieval directly into the reference mobile app (RNFS-cached JSON, refreshed on launch and on demand — no model re-download required)
• Extended the news RAG with AI/tech feeds (TechCrunch AI, The Verge AI, MIT Technology Review), added Atom feed parsing, strengthened the grounding instruction — result honestly documented as mixed (see above), not oversold
• Rendered both architecture diagrams as actual images after finding Hugging Face's model card renderer doesn't support Mermaid (unlike GitHub)
• Model cards cleaned up: no base-model linkage, no license link naming anything specific — generic Apache 2.0 declaration only
• Added a creator-identity system prompt instruction, tested across phrasings
• Added the news RAG pipeline above — retrieval, not retraining; weights untouched
• Fixed GGUF export: two silent bugs in the MLX→GGUF conversion path (a conv1d weight axis-order mismatch, an RMSNorm weight offset convention mismatch) produced a file that loaded without error but generated complete garbage. Fixed and verified; working GGUF published to a separate repo.

2026-09-07

• Added a general-knowledge fine-tuning round (88 spot-checked geography/science/history examples, distilled from a larger local teacher) — verified genuine but partial accuracy improvement, documented honestly including facts that stayed wrong (see Limitations)
• Documented the tool/function-calling limitation
• Rebalanced the xGAIR intent-parsing training data (32 → 69 examples) after finding an earlier small "general knowledge about xGAIR" set was unreliable and caused hallucination; dropped it, kept the structured intent-parsing task that the data scale actually supports
• Initial release: personal-history + general-QA + distillation training pipeline, first published model card

License

Apache 2.0. IRx-1 is a derivative fine-tuned model — full Apache 2.0 terms apply as with any work under this license.