Nano 1.3-Fast — GGUF

GGUF versions of LGTEAMai/Nano-1.3-Fast, a compact bilingual (English and Russian) chat model with 197 million parameters, for llama.cpp, Ollama, LM Studio and other GGUF runtimes. Released 9 November 2026.

The model is tiny, so even the F16 file is only half a gigabyte and runs comfortably on a laptop CPU, a Raspberry Pi 5 or a phone. Read the model card of the original repository for the capabilities and limitations — in short, it is good for simple short chat and unreliable for facts and arithmetic.

Files

*Perplexity on a fixed 311 KB sample of English and Russian text (Wikipedia and web), 100 windows of 512 tokens, measured with llama-perplexity (llama.cpp commit a4d4f2f). Lower is better; only the differences between files matter. For a model this small, deeper quantization costs more quality than for large models, so prefer Q8_0 unless you have a size limit; Q8_0 is only about 60 MB larger than Q4_K_M.

All files were produced with llama.cpp's own convert_hf_to_gguf.py and llama-quantize from the safetensors release (llama.cpp commit a4d4f2f). Verification: the token ids produced by llama.cpp are identical to the Hugging Face tokenizer on 12 English, Russian, code, number and special-token test strings, and llama-server --jinja builds the exact training chat format from the template embedded in the file and stops cleanly at <|end|>.

Chat format

The chat template is embedded in the files (Jinja), producing:

<|system|>
{system}
<|end|>
<|user|>
{message}
<|end|>
<|assistant|>

The model ends its reply with <|end|>. Always give it the recommended system prompt, it behaves noticeably worse without one:

You are Nano 1.3-Fast-Instruct, a compact AI assistant developed by LGTEAMai. Answer helpfully and concisely, in the language of the user.

Recommended sampling: temperature 0.6, top-k 40, top-p 0.9, repetition penalty 1.1. Context length: 1,024 tokens.

Usage

llama.cpp

llama-server -m Nano-1.3-Fast-Q8_0.gguf --jinja -c 1024 --temp 0.6 --top-k 40 --top-p 0.9 --repeat-penalty 1.1

then open http://localhost:8080 or use the OpenAI-compatible API (/v1/chat/completions) and send the system prompt above as the first message. The chat template stored in the GGUF is applied automatically.

Interactive chat in the terminal:

llama-cli -m Nano-1.3-Fast-Q8_0.gguf -c 1024 --temp 0.6 --top-k 40 --top-p 0.9 --repeat-penalty 1.1   -sys "You are Nano 1.3-Fast-Instruct, a compact AI assistant developed by LGTEAMai. Answer helpfully and concisely, in the language of the user."

One-shot completion with the chat format written by hand (llama-completion is the non-interactive tool in current llama.cpp):

llama-completion -m Nano-1.3-Fast-Q8_0.gguf -no-cnv --temp 0.6 --top-k 40 --top-p 0.9 --repeat-penalty 1.1 -n 200   -p $'<|system|>
You are Nano 1.3-Fast-Instruct, a compact AI assistant developed by LGTEAMai. Answer helpfully and concisely, in the language of the user.
<|end|>
<|user|>
Who created you?
<|end|>
<|assistant|>
'

Ollama — create a Modelfile:

FROM ./Nano-1.3-Fast-Q8_0.gguf
TEMPLATE """{{- if .System }}<|system|>
{{ .System }}
<|end|>
{{ end }}{{- range .Messages }}{{- if eq .Role "user" }}<|user|>
{{ .Content }}
<|end|>
{{ else if eq .Role "assistant" }}<|assistant|>
{{ .Content }}<|end|>
{{ end }}{{- end }}<|assistant|>
"""
SYSTEM """You are Nano 1.3-Fast-Instruct, a compact AI assistant developed by LGTEAMai. Answer helpfully and concisely, in the language of the user."""
PARAMETER stop "<|end|>"
PARAMETER temperature 0.6
PARAMETER top_k 40
PARAMETER top_p 0.9
PARAMETER repeat_penalty 1.1
PARAMETER num_ctx 1024

ollama create nano-1.3-fast -f Modelfile
ollama run nano-1.3-fast

LM Studio — download a file (Q8_0 recommended) into LM Studio's models folder, select it, paste the system prompt above and set the sampling values listed above.

About the model

Nano 1.3-Fast is developed by LGTEAMai, trained from scratch, and is part of the "AI makes AI" research project. It is released under the Apache 2.0 license. Full details — training data, evaluation, intended use and limitations — are on the original model card.