Instructions to use LGTEAMai/Nano-1.3-Plus with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use LGTEAMai/Nano-1.3-Plus with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="LGTEAMai/Nano-1.3-Plus") messages = [ {"role": "user", "content": "Who are you?"}, ] pipe(messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("LGTEAMai/Nano-1.3-Plus") model = AutoModelForCausalLM.from_pretrained("LGTEAMai/Nano-1.3-Plus", device_map="auto") messages = [ {"role": "user", "content": "Who are you?"}, ] inputs = tokenizer.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use LGTEAMai/Nano-1.3-Plus with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "LGTEAMai/Nano-1.3-Plus" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "LGTEAMai/Nano-1.3-Plus", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/LGTEAMai/Nano-1.3-Plus
- SGLang
How to use LGTEAMai/Nano-1.3-Plus with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "LGTEAMai/Nano-1.3-Plus" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "LGTEAMai/Nano-1.3-Plus", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "LGTEAMai/Nano-1.3-Plus" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "LGTEAMai/Nano-1.3-Plus", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Docker Model Runner
How to use LGTEAMai/Nano-1.3-Plus with Docker Model Runner:
docker model run hf.co/LGTEAMai/Nano-1.3-Plus
Nano 1.3-Plus
Released 15 November 2026. Model cards may be updated from time to time.
Nano 1.3-Plus 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 โ no custom code. It introduces itself as Nano 1.3-Plus.
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-Plus-GGUF.
Model Information
Description. An instruction-tuned version of a Nano 1.3 pretrained model that was trained for 12.6 billion tokens (about 64 tokens per parameter) and extended to a 4,096-token context. Part of LGTEAMai's "AI makes AI" research, which studies whether current AI models can help build their smaller successors under human supervision. Compared with Nano 1.3-Fast it has a four times longer context, answers short everyday questions (greetings, thanks, definitions, small arithmetic) in a fixed, short style, writes basic programs such as "Hello, World!" and draws simple shapes as SVG.
Model dependencies. Trained from scratch by LGTEAMai (Nano 1.2 โ Nano 1.3 โ this model). It is not a fine-tune of another company's model.
Inputs. Text. Context length: 4,096 tokens (RoPE base 50,000).
Outputs. Text. The default generation limit is 2,048 new tokens; see the limitations for the quality of long answers.
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.
Knowledge cutoff. Encyclopedic knowledge in the training mix goes up to September 2026; web text is older. The model is trained to say that its information runs to September 2026, but it does so only part of the time (see the limitations).
Quick start
from transformers import AutoTokenizer, AutoModelForCausalLM
repo = "LGTEAMai/Nano-1.3-Plus"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo)
messages = [
{"role": "system", "content": "You are Nano 1.3-Plus, 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.
- Ask one thing at a time. Short answers work best; long answers can start to repeat (see the limitations).
- 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. Publicly available text and code in English and Russian: web, educational, encyclopedic, mathematical and source-code text, including a small share of vector-graphics (SVG) code, including encyclopedic text up to September 2026.
Instruction tuning. About 64,000 conversations: English and Russian chats from publicly available data, plus examples written by LGTEAMai โ the model's identity, everyday small talk, short factual answers, definitions, arithmetic, questions about the date and the knowledge cutoff, basic programming and SVG drawings. 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 3,072 tokens were cut at the last complete turn. The loss was computed on the assistant's tokens only.
Implementation and Sustainability
Hardware. Pretraining stages on NVIDIA H100, H200 and T4 GPUs; instruction tuning on one H100 in about 15 minutes (1,994 steps, two passes over the data).
Software. PyTorch; training code written in-house.
Distribution
Open weights under the Apache 2.0 license on Hugging Face:
LGTEAMai/Nano-1.3-Plusโ safetensors (this repository)LGTEAMai/Nano-1.3-Plus-GGUFโ 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 v1.1 and Nano-Micro-Bench v1.1, two small judge-free benchmarks published by LGTEAMai. Random-guess accuracy is shown for reference; differences of 1โ2 points are within measurement error.
Results.
| Benchmark | Metric | Random guess | Nano 1.3-Plus | Nano 1.3-Fast | Nano 1.2 Optimal |
|---|---|---|---|---|---|
| ARC-Easy | accuracy | 25.0% | โ 40% | 39.9% | 39.0% |
| ARC-Challenge | normalized accuracy | 25.0% | 23.3% | 22.4% | 21.9% |
| HellaSwag | normalized accuracy | 25.0% | 29.9% | 29.8% | 30.0% |
| PIQA | accuracy | 50.0% | 60.6% | 61.1% | 61.2% |
| Winogrande | accuracy | 50.0% | 48.9% | 50.4% | 52.2% |
| COPA | accuracy | 50.0% | 59.0% | 60.0% | 59.0% |
| SciQ | accuracy | 25.0% | 68.3% | 68.3% | 65.8% |
| Nano-Chat-Bench v1.1 | composite | 0% | 35.5% | 33.0% | 17.1% |
| Nano-Micro-Bench v1.1 | composite | 0% | 79.4% | 78.6% | 46.2% |
The standard benchmarks are knowledge and commonsense tests, not chat tests. Instruction tuning keeps the knowledge level of the pretrained model: Nano 1.3-Plus and Nano 1.3-Fast are equal on all of them within measurement error (standard errors are 0.5โ1.5 points, 4.9 for COPA). In Nano-Chat-Bench v1.1 the two Nano 1.3 models are also almost tied: Plus stops by itself and avoids repeating itself more often (Hygiene 94.2% against 82.7%), Fast follows slightly more of the simple instructions (5 of 24 against 3).
Comparison with public chat models (Nano-Chat-Bench v1.1, same items and settings).
| Model | Parameters | Pretraining tokens (as reported) | Facts | Instructions | Hygiene | Quality | NCB |
|---|---|---|---|---|---|---|---|
| Gemma 3 1B IT | 1.0B | โ 2T | 73.3% | 83.3% | 95.5% | 91.1% | 86.2% |
| Qwen2.5-0.5B-Instruct | 0.49B | โ 18T | 71.7% | 54.2% | 96.7% | 73.3% | 74.9% |
| SmolLM2-360M-Instruct | 0.362B | โ 4T | 51.7% | 41.7% | 86.7% | 43.3% | 55.9% |
| SmolLM2-135M-Instruct | 0.135B | โ 2T | 30.0% | 25.0% | 78.7% | 26.7% | 40.2% |
| Nano 1.3-Plus | 0.197B | 12.6B | 3.3% | 12.5% | 94.2% | 28.9% | 35.5% |
| Nano 1.3-Fast | 0.197B | 12B | 3.3% | 20.8% | 82.7% | 24.4% | 33.0% |
| GPT-2 small (base model, few-shot prompt, for reference) | 0.124B | โ 10B (estimated) | 5.0% | 4.2% | 75.6% | 0.0% | 21.0% |
Nano 1.3-Plus is clearly behind public chat models of similar or larger size, which were trained on roughly 160 to 1,400 times more text (2โ18 trillion tokens against 12.6 billion). Nano-Chat-Bench v1.1 has harder factual questions than v1.0 and none of them appears in the fine-tuning data; Nano 1.3-Plus answers 2 of the 60. Where it holds up is chat "hygiene": it stops cleanly, always answers in the language of the question (SmolLM2-360M and SmolLM2-135M answered in the wrong language in a third and 40% of the open prompts) and does not repeat itself. Treat it as a small, transparent research model, not as a replacement for these.
Language modelling (Nano-Micro-Bench v1.1). Used as a plain text-completion model, Nano 1.3-Plus scores 91.3% in English and 67.5% in Russian: it writes grammatical, mostly connected text, and in Russian it is weaker at choosing the word that fits the context.
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. In our probes it answered "2 + 2" correctly but "7 + 8" and "6 ร 7" incorrectly, and it did not know the capital of Peru or the author of The Starry Night. Never rely on it for factual answers.
- SVG drawings are well-formed but often do not show what was asked. In our probes every answer was valid SVG, and requests for a circle, a square or a triangle were drawn as such in 22 of 24 attempts; requests for objects (a house, a tree, a cat, a smiley) produce valid but unrelated shapes.
- Long answers. The default output limit is 2,048 tokens and the context is 4,096 tokens, but long answers are not reliable: in our tests the pretrained model began to loop after roughly 700โ1,000 tokens, and most fine-tuning answers are short. Treat long outputs as experimental.
- The date and the knowledge cutoff. It is trained to answer "what year is it?" with "I cannot tell the exact year, but according to my data it is September 2026". In our probes it did so in 9 of 12 attempts for the three Russian "what year is it" wordings, in 1 of 8 for "what is the date" and "what month is it", and in none of 4 for the English question; the other answers only say that it has no calendar.
- Weak reasoning and instruction following. It handles single-turn, simple requests best; long, multi-step or strictly formatted requests usually fail. In Nano-Chat-Bench it passes 3 of 24 rule-checked instructions, and it answers "Say hello in one short sentence" with a Python "Hello, World!" snippet.
- Identity answers can leak into unrelated questions, especially for very short prompts, and some Russian answers still repeat template phrases about the model itself.
- No memory, no internet, no tools.
- Russian is weaker than English.
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.
Contact
- Downloads last month
- 53
Model tree for LGTEAMai/Nano-1.3-Plus
Evaluation results
- accuracy (0-shot) on ARC-Easytest set self-reported39.940
- normalized accuracy (0-shot) on ARC-Challengetest set self-reported23.290
- normalized accuracy (0-shot) on HellaSwagvalidation set self-reported29.900
- accuracy (0-shot) on PIQAvalidation set self-reported60.610
- accuracy (0-shot) on SciQtest set self-reported68.300