Nyqo n1

Nyqo n1 is a compact, general-purpose language model with 497M parameters, built on LiquidAI/LFM2.5-350M. It expands the base model from 16 to 24 layers and trains only the eight added layers, while the original weights stay frozen.

Post-training combines supervised fine-tuning on verified answers with distillation from the base model and an earlier Nyqo checkpoint. Corrected and external reference answers receive supervised loss only. The selected checkpoint saw 800 examples from a fixed set of 1,600; the training corpus is private.

Post-trained on one NVIDIA RTX 3060 (12 GB), with US$19.23 in reported cloud instance rental costs.

Training method · Evaluation record

Property	Value
Parameters	497,119,488 total; 142,635,520 trained
Architecture	LFM2 hybrid convolution–attention, 24 layers
Language	English (training and evaluation)
Format	Full model in Safetensors
Checkpoint identity	File manifest and hashes

Evaluation

Scores as a text table

With a limited compute budget, we evaluated fixed subsets of ARC-Easy (594 tasks), IFEval (210 tasks), and MBPP+ (100 tasks). These are subset scores, not full-benchmark leaderboard results. Internal eval is our own evaluation set of 204 tasks.

The additional comparison scores are reported by the author as descriptive references; exact checkpoint IDs, revisions, and per-model inference records are not included in the release evidence.

For n1 and the base model, ARC-Easy is scored by normalized answer-choice likelihood, IFEval by strict prompt-level accuracy, and MBPP+ requires passing both the base and additional tests from one greedy response.

n1 improves on the ARC-Easy and MBPP+ subsets, while its IFEval score is lower. The paired 95% confidence intervals for the ARC-Easy and MBPP+ gains lie above zero; the interval for the IFEval difference includes zero. These analyses cover n1 versus the base model; paired intervals for the additional comparison models are not reported.

MBPP+ format compliance rose from 45 to 97 of 100 responses, which may contribute to the gain; that contribution has not been isolated.

On the internal evaluation set, n1 scores 56/204, versus 45/204 for the base model (+24.44% relative; +5.39 percentage points). This meets the project's acceptance criteria. The earlier internal bootstrap interval included zero; the updated aggregate summary has no recomputed paired interval. Raw internal item records were not recovered in the checked backups, so the current total cannot be independently regraded from them.

Full protocols, confidence intervals, historical results, and artifact availability are in the evaluation record.

Quick start

The checkpoint files are included in this repository. File integrity and local CPU loading and a short greedy generation have been checked. The CUDA example below has not yet been validated in a fresh GPU environment. It requires a CUDA GPU with BF16 support. The recorded environment used Python 3.12, PyTorch 2.11.0, and Transformers 5.8.1; pinned dependencies are provided in the GitHub repository. Release checks.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Hxtomi/nyqo"

tokenizer = AutoTokenizer.from_pretrained(model_id)
messages = [{"role": "user", "content": "Explain gradient descent in simple terms."}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt", return_dict=True
).to("cuda")

max_new_tokens = 768
if inputs["input_ids"].shape[1] + max_new_tokens > 2048:
    raise ValueError("Prompt and response must fit within 2,048 tokens.")

model = AutoModelForCausalLM.from_pretrained(
    model_id, dtype=torch.bfloat16, attn_implementation="sdpa"
).to("cuda").eval()

with torch.inference_mode():
    output = model.generate(
        **inputs,
        do_sample=False,
        num_beams=1,
        max_new_tokens=max_new_tokens,
        use_cache=True,
        eos_token_id=tokenizer.eos_token_id,
        pad_token_id=tokenizer.pad_token_id,
    )

response = output[0, inputs["input_ids"].shape[1]:]
print(tokenizer.decode(response, skip_special_tokens=True))

This uses the native chat template, BF16/SDPA, and greedy decoding, reserving 768 new tokens within a total budget of 2,048. That is the evaluated generation profile, not the maximum supported context length; IFEval used a separate 4,096-token window.

For file verification, multi-turn JSON input, and extraction of the final answer after a leading <think>…</think> block, use the generation CLI and usage guide.

Limitations

• Results cover one post-training seed and one selected checkpoint. They do not establish robustness across training runs or performance on the full public benchmarks.
• Retention is incomplete: the earlier item-level analysis found 19 lost base-model successes. An aggregate gain does not imply improvement on every task or domain.
• The experiments do not isolate the effects of expansion, answer selection, and distillation. There is no same-size fine-tuning control matched to the final 1,600-example stage, no ablation without distillation, and no measured inference speed. Earlier exploratory runs are described in the architecture history.
• Overlap checks covered training inputs using exact matching and bounded semantic retrieval. They were not exhaustive; training targets were not compared, and overlap with upstream pretraining is unknown.
• Other languages, longer contexts, tool use, long-horizon agent behavior, production reliability, and safety have not been evaluated.

License and credits

Model files are distributed under the LFM 1.0 License, inherited from LiquidAI/LFM2.5-350M. See NOTICE for modifications and attribution. Nyqo-authored source code and documentation are under Apache 2.0. The training corpus and closed evaluation bank are not released.

Nyqo team: Artur Markov-Tsoy · Daniel Ayupov.

Built on Liquid AI's LFM2.5-350M. Evaluation uses ARC, IFEval, and EvalPlus (MBPP+); their materials retain their own licenses. Citation metadata.