Instructions to use q-project/N-Q-26M with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use q-project/N-Q-26M with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="q-project/N-Q-26M", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("q-project/N-Q-26M", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use q-project/N-Q-26M with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "q-project/N-Q-26M"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "q-project/N-Q-26M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/q-project/N-Q-26M

SGLang

How to use q-project/N-Q-26M 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 "q-project/N-Q-26M" \
    --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": "q-project/N-Q-26M",
		"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 "q-project/N-Q-26M" \
        --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": "q-project/N-Q-26M",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use q-project/N-Q-26M with Docker Model Runner:
```
docker model run hf.co/q-project/N-Q-26M
```

N-Q-26M

A 26M-class Russian/English base model — mini Qwen3.8-Flash-Next, pretrained from scratch on one V100

This is a research release, not a product. A plain text-continuation model: grammatical Russian and English, confident tone, invented facts. For dialogue and exact tool calls use N-Q-26M-Chat.

Model summary

N-Q-26M is the base model of the smallest rung of N-Q, the Q Project's MiniMind-style ladder of small RU/EN models trained from scratch on a single 16GB V100. Architecture: a mini Qwen3.8-Flash-Next — Gated DeltaNet linear attention interleaved 3:1 with gated full attention, a 4-stream Gated Residual, and a hashed n-gram embedding — chosen over the ternary Q-164M design by a controlled ablation (see the chat model card for the table).

29.7M backbone parameters + 16.8M n-gram table (46.5M stored), 8,256-token RU+EN byte-level BPE, pretrained on 2.0B tokens in 6.6 hours (30,517 steps, 64×1024 tokens/step, WSD schedule, Muon + AdamW, torch.compile, 85k tokens/s). 5% of rows were dialogues and <CALC> circuit rows from step 1 (the Q-164M recipe), so even the base model knows the chat and tool formats.

Results (final checkpoint, step 30,517)

held-out set (never trained on)	loss	perplexity (this tokenizer)
FineWeb-Edu (EN)	2.381	10.8
Cultura-Ru-Edu (RU)	2.651	14.2
Russian Wikipedia	2.323	10.2

Perplexities are per token of this 8,256-token tokenizer and are not comparable to models with other tokenizers. For scale: the same architecture after 0.5B tokens (ablation run) reached a mean loss of 2.616; 2.0B tokens bring it to 2.452.

Samples (temperature 0.7):

Москва — столица Московской области. Парк расположен в 35 км от Москвы… В годы Великой Отечественной войны город подвергся нападению…

Once upon a time, in ancient times, a small town called Polar, was located on the town of Benedict in Poland. In the 14th century…

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
tok = AutoTokenizer.from_pretrained("q-project/N-Q-26M", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("q-project/N-Q-26M", trust_remote_code=True)
ids = tok("Москва — столица", return_tensors="pt")
out = model.generate(ids, max_new_tokens=80, temperature=0.7, top_p=0.9, stop_ids=(2,))
print(tok.decode(out[0], skip_special_tokens=True))

The model has its own generate() (recurrent Gated DeltaNet state + KV cache). Pass tokenizer=tok to switch on the exact <CALC> circuits if you prompt it in the chat format.

Packed single-file format (`model.qpack`)

QPACK1 container (same format as Q-164M / Q-Omni): weights, config, tokenizer and a self-test in one mmap-friendly file, 89 MiB vs 177 MiB safetensors. A folder or repo with model.qpack and no *.safetensors loads through AutoModelForCausalLM.from_pretrained(..., trust_remote_code=True); python qpack.py --selftest model.qpack prints SELFTEST 5/5. Matrices fp16, 1-D tensors fp32, n-gram hash tables exact int64.

Architecture

8 layers, pattern LLLA ×2: Gated DeltaNet (3 key / 6 value heads, dim 64, conv k=4) ×3, then gated attention (8 query / 2 KV heads, dim 64, zero-centred QK-norm, partial RoPE 25% θ=1e6, sigmoid output gate).
hidden 512, SwiGLU 512→704→512, Gated Residual with 4 residual streams, zero-centred RMSNorm, untied output head.
Hashed 3-gram embedding (PLE) after layer 2: 16.8M parameters in prime-sized tables.
Token layout: ids 0–8191 text (0 <pad>, 1 <bos>, 2 <eos>), 8192–8255 control tokens (<USER> <MODEL> <EOT> <THINK> <ETHINK> <CALC> <EQ> <ECALC> <NEED> <QUOTE> <SYSTEM>, reserved <R11>…<R63>).

Limitations

Facts are frequently wrong; thin general knowledge; repetitive under greedy decoding. No safety tuning.
Base model: no instruction following beyond the 5% dialogue rows it saw in pretraining.
Research and education only.

License and data terms

Code and weights: Apache-2.0 (see LICENSE). Training data keeps its own terms — check before commercial use: FineWeb-Edu (ODC-By 1.0), Cultura-Ru-Edu, Russian Wikipedia 20231101.ru (CC BY-SA / GFDL); dialogue/circuit rows from SmolTalk, GrandMaster-PRO-MAX, saiga_scored and our own <CALC> corpus.

Downloads last month: -

Safetensors

Model size

46.5M params

Tensor type

I64

F32

Model tree for q-project/N-Q-26M

Finetunes

1 model

Datasets used to train q-project/N-Q-26M

Collection including q-project/N-Q-26M

N-Q

Collection

3 items • Updated about 4 hours ago • 1