# Mathematical Reasoning

OpenR1-Distill-7B is a post-trained version of Qwen2.5-Math-7B on the Mixture-of-Thoughts dataset, designed to teach language models step-by-step reasoning.

Thinkless is a large language model trained via reinforcement learning, capable of adaptively selecting short or long-chain reasoning modes, significantly reducing inference computational costs.

The Thinkless framework is a learnable framework that enables large models to adaptively choose between short reasoning or long-chain reasoning based on task complexity and their own capabilities.

An instruction-tuned model based on the Qwen3-14B architecture, specializing in mathematical reasoning, code generation, and factual accuracy

E1-Math-1.5B is a language model fine-tuned based on DeepSeek-R1-Distilled-Qwen-1.5B, supporting elastic reasoning and the GRPO method, suitable for budget-constrained deduction scenarios.

OLMo 2 1B Instruct Edition is a post-training variant of the OLMo-2-0425-1B-RLVR1 model, optimized through supervised fine-tuning, DPO training, and RLVR training to achieve state-of-the-art performance across multiple tasks.

Phi-4-mini-reasoning is a lightweight open-source model focused on mathematical reasoning tasks, supporting a context length of 128K tokens.

OLMo 2 1B is a post-training variant of the allenai/OLMo-2-0425-1B-RLVR1 model, undergoing supervised fine-tuning, DPO training, and RLVR training, aiming to achieve state-of-the-art performance across multiple tasks.

OpenMath-Nemotron-32B is a mathematical reasoning model created by fine-tuning Qwen2.5-32B on the OpenMathReasoning dataset, achieving state-of-the-art results on multiple mathematical benchmarks.

A mathematical reasoning model fine-tuned based on Qwen2.5-14B, achieving first place in the AIMO-2 Kaggle competition

OpenMath-Nemotron-14B is a mathematical reasoning model fine-tuned on the OpenMathReasoning dataset based on Qwen2.5-14B, achieving state-of-the-art results on multiple mathematical benchmarks.

A next-generation language model based on the Qwen 2.5 14B architecture, specifically optimized for mathematical reasoning, programming, and general logical tasks.

Phi-4 Reasoning is a cutting-edge open-weight reasoning model based on Phi-4, fine-tuned with supervised chain-of-thought trajectory datasets and trained via reinforcement learning, specializing in mathematics, science, and programming skills.

ReasonFlux-F1-7B is a hierarchical large language model reasoning model based on thinking template expansion, fine-tuned through template-enhanced reasoning trajectories, demonstrating outstanding performance in multiple reasoning tasks.

14B-parameter-scale mathematical reasoning large language model supporting chain-of-thought reasoning

ThinkEdit is a lightweight weight editing method that identifies and edits a small number of attention heads to mitigate the issue of large language models generating overly short reasoning chains in inference tasks, thereby improving reasoning accuracy.

A model fine-tuned based on Google Gemma-3-4B-IT, specifically designed for solving various high school-level problems

The EXAONE Deep series models excel in reasoning tasks such as mathematics and programming. This model is the 4-bit AWQ quantized version with 2.4 billion parameters

QwQ-Math-IO-500M is a 500M-parameter language model focused on mathematical reasoning and input-output processing, offering quantized versions in GGUF format.

SoT_DistilBERT is a classification model fine-tuned based on DistilBERT, designed to select the optimal reasoning paradigm for a given query according to the Sketch-of-Thought (SoT) framework.

YuLan-Mini-Instruct is a compact yet powerful 2.4-billion-parameter text generation model, specializing in mathematical and code reasoning tasks with support for both English and Chinese.

GSM8K is a dataset containing 8.5K high-quality elementary school math problems, designed to evaluate models' mathematical reasoning abilities.

AceMath-7B-Instruct is a specialized instruction model for mathematical reasoning developed by NVIDIA, based on an improved Qwen architecture, excelling at solving English math problems through chain-of-thought (CoT) reasoning.

NanoLM-1B-Instruct-v2 is a 1B-parameter small model fine-tuned on over 4 million high-quality instruction datasets, dedicated to exploring the potential of small models.

Mathstral 7B is a specialized model for mathematical and scientific tasks, developed based on Mistral 7B, excelling in mathematical reasoning and scientific computing tasks.

InternLM-Math-Plus is the state-of-the-art bilingual open-source mathematical reasoning large language model, equipped with solving, proving, verifying, and enhancing capabilities.

Llama3-70B-Chinese-Chat is a large bilingual language model developed based on Meta-Llama-3-70B-Instruct, specifically designed for Chinese and English users, with capabilities such as role-playing, tool usage, and mathematical computation.

Granite-20B-Code-Instruct-8K is a 20-billion-parameter code generation model, fine-tuned from Granite-20B-Code-Base-8K, specializing in code generation and instruction-following capabilities.

Snowflake Arctic is a large language model developed by Snowflake AI Research Team, featuring a dense Mixture of Experts (MoE) architecture with 480 billion parameters, specifically designed for efficient text and code generation.

Arctic is a dense Mixture of Experts (MoE) architecture large language model developed by the Snowflake AI Research team, with 480 billion parameters, open-sourced under the Apache-2.0 license.

Neuro 4 Math-7b is a specialized mathematical model created by merging multiple 7B-parameter models using the LazyMerge toolkit, focusing on math-related tasks.

Neural-4-QA-7b is a large language model based on the merging of multiple 7B-parameter models, focusing on Q&A tasks and optimized using the DARE-TIES merging method.

The OpenMath model solves mathematical problems by combining textual reasoning with Python interpreter-executed code blocks, fine-tuned based on Mistral-7B-v0.1

DeepSeekMath is a large language model specialized in solving mathematical problems, capable of handling complex mathematical reasoning and computation tasks.

DeepSeekMath is a model specialized in the field of mathematics, supporting commercial use under the MIT License.

DeepSeekMath is a large language model specialized in solving mathematical problems, featuring powerful mathematical reasoning and computational capabilities.

Stealth-v1.3 is an open-source large language model developed by Jan, supporting offline operation on local devices to ensure user privacy.

WizardMath-7B-V1.1 is a state-of-the-art 7B mathematical large language model trained on Mistral-7B, excelling on GSM8k and MATH datasets.

ToRA-7B is a 7B-parameter tool-integrated reasoning agent, specifically designed for solving mathematical problems by combining natural language reasoning with external tool usage.

Llemma 34B is a language model specialized in the field of mathematics, initialized based on the weights of Code Llama 34B and trained on 50 billion tokens from the Proof-Pile-2 dataset.