🍡 Mochi

Mochi is a math-reasoning fine-tune of GLM-4.7-Flash, trained on the Open Math Reasoning (mini) dataset — the same chain-of-thought data used in the winning submission to the AI Mathematical Olympiad Progress Prize 2 (AIMO-2) on Kaggle.

The goal of this fine-tune is to sharpen GLM-4.7-Flash's step-by-step mathematical reasoning while keeping the small, fast footprint of the Flash base model.

Looking for a quantized/local version? See mochi-gguf for GGUF builds you can run with llama.cpp, Ollama, or LM Studio.

Model Details

Base model: GLM-4.7-Flash
Fine-tuning data: unsloth/OpenMathReasoning-mini — chain-of-thought math solutions distilled from the larger OpenMathReasoning dataset, which underpinned NVIDIA's AIMO-2-winning system
Objective: Improve multi-step mathematical reasoning and answer accuracy on olympiad-style problems
Language(s): English
License: apache-2.0 (inherits from GLM-4.7-Flash — please verify against the base model's license before commercial use)

Intended Use

Mochi is intended for:

Step-by-step mathematical problem solving (algebra, number theory, combinatorics, olympiad-style questions)
Research into chain-of-thought fine-tuning on small/flash-class models
Local or low-latency deployments that still need decent math reasoning

It is not intended as a general-purpose reasoning or safety-critical decision-making tool. As with any LLM, verify important calculations independently.

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "artindnr/mochi"

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

messages = [
    {"role": "user", "content": "If x^2 - 5x + 6 = 0, what are the values of x?"}
]

inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt"
).to(model.device)

outputs = model.generate(inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

You can also load Mochi with Unsloth for faster inference and further fine-tuning:

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="artindnr/mochi",
    max_seq_length=4096,
    load_in_4bit=True,
)

Training

Mochi was fine-tuned starting from GLM-4.7-Flash on the CoT split of unsloth/OpenMathReasoning-mini, using Unsloth for efficient LoRA/QLoRA training.


Base model	GLM-4.7-Flash
Dataset	unsloth/OpenMathReasoning-mini (CoT split)
Task	Supervised fine-tuning (chain-of-thought math)
Framework	Unsloth

Limitations

Fine-tuned specifically for math reasoning; general chat/instruction-following ability may drift from the base model.
Trained on a "mini" subset of OpenMathReasoning, so coverage of problem types is narrower than the full dataset.
Like all LLMs, it can produce confidently incorrect derivations — always double-check final answers on problems that matter.

Citation

If you use this model, please also credit the underlying dataset and competition it draws from:

@misc{openmathreasoning,
  title  = {OpenMathReasoning},
  author = {NVIDIA},
  year   = {2025},
  url    = {https://huggingface.co/datasets/nvidia/OpenMathReasoning}
}