RusNanoLM-800M: an ultra-compact LLM with strong Russian performance
2 min read
RusNanoLM-800M is an open-source, 800M-parameter bilingual language model built for Russian and English. Based on the hybrid Qwen3.5 architecture with linear attention (Gated Delta Rule), it delivers standout scores on the MERA benchmark (an evaluation suite for Russian language models). The model outperforms its base version and beats much larger LLMs on arithmetic reasoning tasks. It is a great fit for developers, researchers, and anyone who needs a fast Russian-capable model to run locally on modest hardware.
Key features
- Hybrid architecture: uses linear attention (Gated Delta Rule) across 24 layers, with full attention on every fourth layer.
- Strong reasoning: outperforms GigaChat-2-10B and Qwen2.5-7B on the RuMultiAR benchmark despite its compact size.
- Bilingual: full out-of-the-box support for both Russian and English.
- Low hardware requirements: takes up roughly 1.7 GB of VRAM in BF16 and runs easily on entry-level GPUs.
What you can do
- Run locally: deploy the model locally using the Python transformers library to test text generation and NLP tasks.
- Use with Ollama: run it directly in the Ollama ecosystem using ready-to-use F16 GGUF files.
- Test reasoning: evaluate the model on logic and arithmetic word problems in Russian.
- Embed in pipelines: integrate it into lightweight services or bots where response speed and minimal resource usage are critical.
Example
import torch
from transformers import AutoModelForImageTextToText, AutoTokenizer
MODEL_ID = "itismax0/RusNanoLM-800M"
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForImageTextToText.from_pretrained(
MODEL_ID,
torch_dtype=torch.bfloat16,
device_map="auto"
)
model.eval()
messages = [
{"role": "user", "content": "Реши задачу: в корзине 5 яблок и 7 груш. Сколько всего фруктов?"}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
outputs = model.generate(
**inputs,
max_new_tokens=512,
do_sample=False,
pad_token_id=tokenizer.eos_token_id
)
response = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)
print(response)Read more
DetailsParameters: 800M (0.8B)
MERA scores: SimpleAR — 83.2%, RuMultiAR — 33.5% (up from 8.1% in the base model)
Hardware requirements: ~1.7 GB VRAM (BF16), runs on 4 GB VRAM GPUs
License: Apache 2.0










Comments