Puro-2B-Base

Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090

Under our fixed 15-benchmark base-model evaluation, one checkpoint in the Puro-2B collection beats Qwen2-1.5B at about $4.4K; the canonical final model goes further, approaching Qwen2.5-1.5B at a measured rental-equivalent accelerator cost of $6,891.

The model weights can also be downloaded from ModelScope.

How Far Can a Poor Lab Go with RTX 5090s?

Puro-2B (普罗-2B) is a 2B-parameter dense causal language model pretrained from scratch on 1.4T tokens. It uses a Qwen3-1.7B-compatible architecture with untied input and output embeddings, blockwise FP8 training, the MuonH optimizer, and a two-phase data recipe. Training ran entirely on consumer-grade NVIDIA RTX 5090 GPUs.

The architecture is based on the Qwen3-1.7B configuration, not on pretrained Qwen weights. Puro-2B starts from random initialization.

<p align="center"> <img src="./assets/cost_quality_pareto.png" width="92%" alt="Puro-2B cost and quality comparison"> </p>

Why Puro-2B?

Puro-2B is intended to make billion-parameter pretraining inspectable and affordable for smaller research groups. The release covers more than the final weights:

• A canonical 2B base model and intermediate or controlled checkpoints: this repo.
• Materialized pretraining data: https://huggingface.co/datasets/thu-pacman/Puro-2B.
• Training implementation: https://github.com/thu-pacman/Puro-Megatron.
• Data-processing implementation: https://github.com/thu-pacman/Kaiyuan-Spark.
• Technical report: https://arxiv.org/abs/2608.27370.

The main recipe combines RTX 5090 infrastructure, blockwise FP8, MuonH with hyperball constraints, proxy-guided data selection, and a curriculum-aware late continuation followed by checkpoint averaging.

<p align="center"> <img src="./assets/pipeline_overview.png" width="96%" alt="Puro-2B end-to-end training pipeline"> </p>

The $5,090 Result, Explained

The collection contains multiple checkpoints with different Phase 2 budgets and recipes. The report's approximately $4.4K result is an observed uniform-recipe checkpoint that already exceeds Qwen2-1.5B on the report's 15-task aggregate. It is not the canonical final checkpoint.

The canonical Puro-2B-Base model is the strongest released endpoint. Its production run used 22,514 measured active-training GPU-hours, corresponding to $6,891 under the report's normalized RTX 5090 rental rate.

These figures are accelerator-only reproduction estimates. They exclude data acquisition and preprocessing, proxy and ablation experiments, failed runs, post-training, evaluation, storage, networking, and research labor. They should not be read as the total cost of developing the project.

<p align="center"> <img src="./assets/cost_saving_factors.png" width="92%" alt="Puro-2B estimated efficiency factors"> </p> <p align="center"> <img src="./assets/renminbi_scaling_law.png" width="78%" alt="Puro Cost Scaling Law across five Phase 2 budgets"> </p>

The scaling-law panel labels points by cumulative reproduction cost. The model catalog below maps those costs to Phase 2 budget fractions. Each fraction applies only to Phase 2 data exposure, while the cost includes the shared Phase 1 run.

Model Details

This is a pretrained base model. It has not been instruction-tuned or preference-aligned and should not be expected to behave like a chat assistant.

Quickstart

Use a Transformers release that supports the Qwen3 configuration:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "thu-pacman/Puro-2B-Base"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

prompt = "The central limit theorem states that"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    do_sample=True,
    temperature=0.7,
    top_p=0.9,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Evaluation

All numbers below come from the same deterministic OpenCompass pipeline in the technical report. The comparison uses pretrained/base checkpoints throughout. Generation tasks use greedy decoding; multiple-choice tasks use fixed token-likelihood ranking. Scores are percentages.

The four math and code tasks are GSM8K, MATH, sanitized-MBPP, and HumanEval. The eleven reasoning and knowledge tasks are MMLU, MMLU-Pro, ARC-Challenge, ARC-Easy, BoolQ, CommonsenseQA, HellaSwag, PIQA, SocialIQA, WinoGrande, and BBH. Each displayed average is an unweighted arithmetic mean.

The Puro Cost Scaling Law fits five single-run Phase 2 uniform-budget points. It is a recipe-specific empirical scale-down relationship, not a universal law. The fit has no uncertainty interval, and the available experiments do not isolate curriculum ordering, constant-LR continuation, and checkpoint averaging as independent causal gains.

Training

Both phases use a sequence length of 4,096, a global batch size of 1,536 sequences, and a micro-batch size of 2. Main Transformer linear-layer GEMMs use blockwise E4M3 FP8; numerically sensitive operations, master weights, and optimizer states remain in BF16 or FP32 as appropriate.

Selected approximately scale-invariant matrix weights are updated by MuonH with hyperball projection and zero weight decay. The remaining parameters use AdamW with weight decay 0.1. The MuonH matrix group applies a 10x multiplier to the shared base learning-rate schedule.

The final model uses an equal-weight parameter average of six checkpoints from the constant-LR branch resumed at optimizer step 218,000:

222100, 222200, 222300, 222400, 222500, 222569

Only model parameters are averaged; optimizer states are not.

Model Catalog

Author-controlled model weights, training code, processing code, and documentation are released under Apache License 2.0 where marked. The materialized dataset is distributed under other because its components retain different upstream terms; see the dataset card's license matrix and notices.

Intended Use and Limitations

Puro-2B is intended for research on pretraining, data recipes, optimization, model scaling, continued pretraining, and downstream adaptation. It can also be used as a compact base model for task-specific post-training.

The model may produce inaccurate, biased, unsafe, offensive, or copyrighted content. Its pretraining data includes web text, code, mathematics, Chinese and English material, synthetic data, and instruction-formatted examples. The release does not claim exhaustive removal of personal information, benchmark contamination, or undesirable content. Evaluate and post-train the model for your domain before deployment, and add application-specific safeguards where people could be affected by its outputs.

License

The Puro-2B model weights are released under the Apache License 2.0. The training data remains subject to the dataset repository's documented upstream licenses and terms.

Citation

Please cite our technical report if you find our work useful:

@misc{luo2026puro2b,
  title        = {Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within {\$}5090},
  author       = {Kairong Luo and Jiarui Cui and Yaorui Yin and Shengqi Chen and
                  Yiming Yang and Linxiang Gao and Yanmohan Wang and Mingzhe Zhang and
                  Kaiyue Wen and Kaifeng Lyu and Wenguang Chen},
  year         = {2026},
  eprint       = {2608.27370},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  url          = {https://arxiv.org/abs/2608.27370}
}

Acknowledgments

We thank Yanfu Investments for providing computational resources. See the technical report for the complete acknowledgments and contributor list.