Nano 1.3-Fast
Released 9 November 2026. Model cards may be updated from time to time.
Nano 1.3-Fast is a compact bilingual (English and Russian) chat model with 197 million parameters, developed by LGTEAMai. It is small enough to run on a laptop CPU or a phone-class device, and it loads with standard transformers — no custom code. It introduces itself as Nano 1.3-Fast-Instruct.
Please read the limitations. This is a very small model. It is good at short, simple chat, but it frequently gets facts and arithmetic wrong and should not be used where correctness matters.
Companion repository with quantized files for llama.cpp, Ollama and LM Studio: LGTEAMai/Nano-1.3-Fast-GGUF.
Model Information
Description. An instruction-tuned version of Nano 1.3 Base, which continues Nano 1.2 pretraining to 12 billion tokens (about 61 tokens per parameter). Part of LGTEAMai's "AI makes AI" research, which studies whether current AI models can help build their smaller successors under human supervision.
Model dependencies. Trained from scratch by LGTEAMai (Nano 1.2 → Nano 1.3 Base → this model). It is not a fine-tune of another company's model.
Inputs. Text. Context length used in training: 1,024 tokens.
Outputs. Text.
Architecture. Decoder-only Transformer in the Llama layout: 16 layers, hidden size 896, 14 attention heads with 2 key/value heads (grouped-query attention), SwiGLU feed-forward (2560), rotary position embeddings, RMSNorm, tied input/output embeddings, 64,007-token byte-level BPE vocabulary (Nano LM-2.1).
Languages. English and Russian. English is somewhat stronger.
Quick start
from transformers import AutoTokenizer, AutoModelForCausalLM
repo = "LGTEAMai/Nano-1.3-Fast"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)
messages = [
{"role": "system", "content": "You are Nano 1.3-Fast-Instruct, a compact AI assistant developed by LGTEAMai. "
"Answer helpfully and concisely, in the language of the user."},
{"role": "user", "content": "Who created you?"},
]
ids = tok.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, max_new_tokens=200) # sampling settings come from generation_config.json
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))
Tips for best results
- Always include the system prompt above. Without it the model drifts more, especially on very short greetings.
- Keep prompts short (the whole conversation must fit in about 1,000 tokens) and ask one thing at a time.
- Use the default sampling (temperature 0.6, top-p 0.9, top-k 40, repetition penalty 1.1). Greedy decoding tends to loop.
Chat format. <|system|>, <|user|>, <|assistant|> turns, each ending with <|end|>; the model stops by emitting <|end|>. The chat template in the tokenizer produces exactly the format used in training.
Model Data
Pretraining (Nano 1.3 stages). Publicly available text in English and Russian: FineWeb-Edu and FineWeb-2 web text, Cosmopedia v2 educational text, Wikipedia (Russian snapshot of 1 September 2026, English snapshot of 2023), FineMath and GitHub code. The earlier Nano 1.2 stages used public web, Wikipedia and code data.
Instruction tuning. About 71,000 conversations: 40,000 English chats from SmolTalk, 27,700 Russian chats from saiga_scored, about 3,600 examples written by LGTEAMai about the model's identity, and about 1,600 short factual answers (capitals, authors, small arithmetic). Conversations that mention other AI models or companies as the assistant's identity were filtered out.
Training data processing. Tokenization with the Nano LM-2.1 tokenizer; conversations longer than 1,024 tokens were cut at the last complete turn.
Implementation and Sustainability
Hardware. Nano 1.3 pretraining stages on 2–4 NVIDIA H100 GPUs (about 5 GPU-hours in total); instruction tuning on one H100 in under 20 minutes.
Software. PyTorch; training code written in-house.
Distribution
Open weights under the Apache 2.0 license on Hugging Face:
LGTEAMai/Nano-1.3-Fast— safetensors (this repository)LGTEAMai/Nano-1.3-Fast-GGUF— GGUF files (F16, Q8_0, Q6_K, Q5_K_M, Q4_K_M)
Evaluation
Approach. Zero-shot, log-likelihood multiple-choice evaluation with the EleutherAI Language Model Evaluation Harness (v0.4.13) on standard benchmarks, and Nano-Chat-Bench v1.0, a small judge-free chat benchmark published by LGTEAMai. Random-guess accuracy is shown for reference; differences of 1–2 points are within measurement error.
Results.
| Benchmark | Metric | Random guess | Nano 1.3-Fast | Nano 1.3 Base | Nano 1.2 Optimal |
|---|---|---|---|---|---|
| ARC-Easy | accuracy | 25.0% | 39.9% | 41.3% | 39.0% |
| ARC-Challenge | normalized accuracy | 25.0% | 22.4% | 23.3% | 21.9% |
| HellaSwag | normalized accuracy | 25.0% | 29.8% | 29.5% | 30.0% |
| PIQA | accuracy | 50.0% | 61.1% | 60.9% | 61.2% |
| Winogrande | accuracy | 50.0% | 50.4% | 49.2% | 52.2% |
| COPA | accuracy | 50.0% | 60.0% | 57.0% | 59.0% |
| SciQ | accuracy | 25.0% | 68.3% | 68.5% | 65.8% |
| Nano-Chat-Bench v1.0 | composite | 0% | 36.1% | 34.0% | 21.5% |
Instruction tuning keeps the knowledge level of the base model (the benchmarks above are knowledge and commonsense tests, not chat tests). In Nano-Chat-Bench, Nano 1.3-Fast stops cleanly, answers in the language of the question and repeats itself less than the base models, but it is not better at facts or following instructions: it passes 2 of 14 simple rule-checked instructions, the same as the base models.
Intended Usage and Limitations
Benefit and intended usage. Research and teaching on small language models; lightweight on-device chat and text drafting in English and Russian; a base for further fine-tuning and for experiments with quantization and distillation.
Known limitations.
- Unreliable facts and arithmetic. It often states wrong facts with confidence and cannot do arithmetic beyond very small numbers. Never rely on it for factual answers.
- Weak reasoning and instruction following. It handles single-turn, simple requests best; long, multi-step or strictly formatted requests usually fail.
- Identity answers can leak into unrelated questions. Because it was trained to introduce itself, it sometimes answers an unrelated question with a description of itself or its abilities, especially for very short prompts.
- Short context (about 1,000 tokens) and no memory, no internet, no tools.
- Russian is weaker than English. The knowledge cutoff has not been formally determined.
Acceptable usage. Not for medical, legal, financial or other safety-critical decisions, and not for generating deceptive or harmful content.
Ethics and Content Safety
Evaluation approach. No formal safety evaluation or red-teaming has been carried out.
Safety policies. The model was not trained with safety alignment. It can produce inaccurate, biased, offensive or otherwise inappropriate text, and it has no reliable refusal behaviour. Applications built on it should add their own filtering and human oversight.
Frontier safety assessment. Given its size and measured capabilities, the model is not expected to approach any frontier capability threshold. No dedicated assessment was conducted.