Greko 2.1 (beta)

Greko is a friendly AI assistant that runs entirely offline on your phone: nothing you type leaves the device, unless you turn on web search. It can see photos, read your documents, and make Word and Excel files. Greko 2.1 is Greko 2 with much better Greek.

New to Greko? Follow the step-by-step install guide (English and Greek), which shows each Android prompt and what to tap.

  • Android app: install greko.apk (Android 13+, arm64, 2 GB of RAM or more). On first launch it offers Download Greko, which fetches the model and its vision encoder once: about 1.3 GB, or 1 GB for Greko 2.1 Lite on phones with less than 2.4 GB of RAM. After that no internet is needed. Phones that already have Greko 2 are offered Greko 2.1 (about 1.1 GB), and phones with the older 4-bit eyes the smaller 3-bit ones (190 MB), so photos need less free memory.
  • Photos and documents (๐Ÿ“Ž): send a photo, or a PDF, Word (.docx) or text file, and ask about it: "summarise this", "what's the deadline in this contract?". Everything is read on the phone. A long document is searched for the parts that match your question, since a phone's model can only take in a few pages at once.
  • Word and Excel files: ask "make me an Excel file withโ€ฆ" or "put this receipt in Excel" with a photo, and Greko's answer comes with a real .docx or .xlsx file to open, share or save.
  • Think mode (๐Ÿ’ก): Greko reasons step by step before answering. Better at maths and logic, but slower.
  • Web search (๐ŸŒ, optional): while the phone is online, a question is searched on DuckDuckGo (or Wikipedia) and Greko answers from the results, listing its sources, with pictures of what you asked about from Wikipedia when it has some. Only the question is sent, to those two sites. It's off until you turn it on.
  • Voice (๐ŸŽค / ๐Ÿ”Š): talk instead of typing, and hear answers read aloud. Speech is recognized and spoken by the phone itself; the voice is never sent anywhere.
  • Updates: once a day the app checks this repository for a newer version and installs it on your OK (only the version file is fetched; nothing about you is sent).
  • Languages: English and Greek; the app follows the phone's language.
  • Made by: George

Files

File What it is
greko-2.1-Q4_K_M.gguf Greko 2.1: Qwen3-VL-2B-Instruct fine-tuned with QLoRA (language model only), 4-bit
greko-2.1-lite.gguf Greko 2.1 Lite for 2 GB phones: the same model at IQ3_M with an importance matrix
mmproj-greko-2-q3.gguf Greko's eyes: the vision encoder and projector, unchanged from Qwen3-VL-2B, at 3 bits (Q3_K, 190 MB). The app uses these, for Greko 2 and 2.1 alike: in tests they read photos as well as the 4-bit and 8-bit ones, a little more slowly
mmproj-greko-2-q4.gguf The same eyes at 4 bits (241 MB), used by the app until 2.7-beta
mmproj-greko-2.gguf The same eyes at 8 bits (445 MB), for PCs
greko.apk The Android app (2.10-beta)
app-version.json What the app reads once a day to offer its own updates
greko-2-Q4_K_M.gguf, greko-2-lite.gguf Greko 2, the previous model, kept for older installs
greko-2.5-beta.apk An earlier app (2.5-beta), in case the current one has a problem on your phone
greko-1.7b-beta-Q4_K_M.gguf, greko-1-beta.apk Greko 1 beta (Qwen3-1.7B, text only)

Training

Greko 2.1 was fine-tuned on 22,423 conversations:

  • 12,000 general conversations from smol-smoltalk
  • 8,557 Greek conversations (chat, instructions and maths) from the Greek part of openeurollm/Dolci-Instruct-SFT-translated, filtered: no code, mostly Greek, no other assistant's name
  • chats written for Greko's persona in English and Greek, hand-written Greek conversations, and multi-turn chats that teach it to answer from the conversation so far
  • step-by-step reasoning examples for Think mode, written by Qwen3-1.7B for questions from the train splits of GSM8K and ARC-Easy, keeping only those that reached the right answer
Qwen3-VL-2B Greko 2 Greko 2.1
Held-out "what's your name" questions naming Greko 4/5 5/5 5/5
Held-out "who made you" questions naming George 0/2 2/2 2/2
GSM8K (100 test questions) 82% 71% 69%
ARC-Easy (200 test questions) 91% 91% 92%
Greek perplexity on 200 held-out Greek chats (lower is better) 1.82 1.63

The GSM8K and ARC-Easy differences between Greko 2 and 2.1 are within noise (ยฑ9 and ยฑ4 points at these sample sizes).

Beta: Greko 2.1's Greek is much better than Greko 2's, but it still gets facts wrong and can ramble in longer Greek answers (summaries, letters). Its maths is below the base model's. Check anything important.

On a PC

llama-mtmd-cli -m greko-2.1-Q4_K_M.gguf --mmproj mmproj-greko-2-q3.gguf --image photo.jpg -p "What's in this photo?"

Greko's system prompt is built into the chat template.

License

Apache 2.0, the same as Qwen3-VL-2B. The Greek training data's license terms are on its dataset page.

Downloads last month
3,452
GGUF
Model size
2B params
Architecture
qwen3
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for geolazagr/greko

Quantized
(99)
this model

Space using geolazagr/greko 1