# Mathematical Reasoning Optimization

Phi-4-reasoning-plus is an open-source reasoning model developed by Microsoft Research, focusing on advanced reasoning capabilities in mathematics, science, and programming fields.

This is a quantized version of Microsoft's Phi-4-reasoning model, optimized using llama.cpp for inference tasks and supporting multiple quantization options.

MiMo-7B-RL is a reinforcement learning model trained based on the MiMo-7B-SFT model, achieving performance comparable to OpenAI o1-mini in mathematical and code reasoning tasks.

OLMo-2-Instruct-Math-32B is a large language model focused on mathematical tasks, released by tngtech.

OpenMath-Nemotron-1.5B is a mathematical reasoning model fine-tuned on the OpenMathReasoning dataset based on Qwen2.5-Math-1.5B, achieving state-of-the-art results on multiple mathematical benchmarks.

Zero-Mistral-24B is an improved text-only model based on Mistral-Small-3.1-24B-Instruct-2503, primarily adapted for Russian and English, with the original visual capabilities removed to focus on text generation tasks.

OpenMath2-Llama3.1-8B is a math-specialized model fine-tuned from the Llama3.1-8B-Base model using the OpenMathInstruct-2 dataset, demonstrating excellent performance across multiple mathematical benchmarks.

Viper-Coder-v1.7-Vsm6 is a large language model based on the Qwen2.5 14B modal architecture, focusing on improving coding efficiency and computational reasoning capabilities, optimizing memory usage, and reducing redundant text generation.

Phi-4-reasoning-plus is an advanced open-weight reasoning model developed by Microsoft Research, optimized through supervised fine-tuning and reinforcement learning based on Phi-4, focusing on advanced reasoning capabilities in mathematics, science, and coding fields.

A text generation model based on google/gemma-3-12b-it, enhanced with the EZO training method to improve Japanese performance

A 24B-parameter large language model optimized for mathematical reasoning and Python programming training, based on the Mistral architecture

OpenRS3-GRPO-ja is a fine-tuned version of the SakanaAI/TinySwallow-1.5B-Instruct model on a Japanese mathematical instruction dataset, trained using the GRPO method, focusing on mathematical reasoning tasks.

Notbad v1.0 Mistral 24B is a model focused on mathematical and Python programming reasoning, based on Mistral-Small-24B-Instruct-2501 and further trained with reinforcement learning.

The EXAONE Deep series models excel in reasoning tasks such as mathematics and programming. The 7.8B version outperforms open-source models of similar scale and even surpasses certain proprietary models.

Based on the OLMo-2-0325-32B-Instruct model, fine-tuned using the Open R1 math dataset on AMD MI300X GPUs, focusing on enhancing mathematical reasoning capabilities

FastCuRL-1.5B-Preview is a slow-thinking reasoning model that employs curriculum-guided iterative extended reinforcement learning, excelling in mathematical reasoning tasks.

A high-performance mathematical reasoning and general knowledge processing model distilled from Qwen2.5-72B through reinforcement learning, excelling in mathematical reasoning and general knowledge tasks.

Gemma-3-4b Reasoning is a Transformer-based language model fine-tuned using the GRPO method, specializing in reasoning task optimization.

A high-performance distilled model optimized for mathematics and general reasoning, refined from Qwen2.5-72B through reinforcement learning

A fine-tuned version based on Qwen/Qwen2.5-7B-Instruct, specifically optimized for advanced reasoning tasks

Designed based on Qwen 2.5 14B modal architecture, enhancing coding efficiency and computational reasoning capabilities

Tiny-R1-32B-Preview is an inference model based on Deepseek-R1-Distill-Qwen-32B, focusing on mathematics, coding, and scientific fields, with performance close to the full version of the R1 model.

A mathematical reasoning model fine-tuned based on Mistral-Small-24B-Instruct-2501, optimized for mathematical reasoning capabilities

An optimized 32B inference model preference version based on Sky-T1-32B-Preview, significantly reducing generation length while maintaining accuracy.

The GGUF format version of the InternLM3-8B-Instruct model, suitable for the llama.cpp framework and supporting multiple quantization versions.

Tülu3 is a new generation of instruction-following model family developed by the Allen Institute for Artificial Intelligence, excelling in standard chat applications and complex problem-solving.

Rho-1 is a language model specialized in mathematics, pretrained with Selective Language Modeling (SLM) method, significantly improving accuracy in solving mathematical problems.

A supervised fine-tuned model based on Smaug-34B, focused on enhancing mathematical and reasoning capabilities, excelling among 34B-scale models.

WestSeverus-7B-DPO-v2 is a WestLake family model trained based on WestSeverus-7B, with training on multiple DPO datasets, demonstrating excellent performance in basic mathematical problems.

Neural-Chat-v3-3 is a 7-billion-parameter large language model developed by Intel based on the Mistral-7B architecture, focusing on mathematical reasoning and text generation tasks. The model is fine-tuned on the MetaMathQA dataset and aligned using Direct Performance Optimization (DPO) method.