# Mathematical reasoning optimization

A static quantization version of UniReason-Qwen3-14B-RL, suitable for text generation and mathematical reasoning research scenarios.

MiMo is a series of 7B parameter models trained from scratch for inference tasks. Through optimized pre-training and post-training strategies, it performs excellently in mathematical and code reasoning tasks.

ALP_DeepScaleR_1.5B_C16K is a model trained using the Adaptive Length Penalty (ALP) method based on the DeepScaleR-1.5B model, which can significantly reduce token usage while maintaining performance.

Multiverse-32B is the first open-source, non-autoregressive model built on Multiverse. It performs excellently in the AIME test and has important academic and application value.

INT4 quantized version of Qwen3-30B-A3B, reducing disk and GPU memory requirements by 75% while maintaining high performance.

Phi-4-mini-reasoning is a lightweight open model built on synthetic data, focusing on high-quality, reasoning-rich data, and further fine-tuned for more advanced mathematical reasoning capabilities.

Phi-4-reasoning is an advanced reasoning model developed by Microsoft, fine-tuned based on Phi-4, focusing on improving reasoning abilities in fields such as mathematics, science, and coding.

Phi-4-mini-reasoning is a lightweight open-source model focused on mathematical reasoning, supporting a context length of 128K tokens and suitable for environments with limited computing resources.

This is a quantized version of the Microsoft Phi - 4 - mini - reasoning model, which is quantized using the llamacpp tool to improve the model's operating efficiency and performance in different hardware environments.

A quantized version of Tongyi Qianwen 3 0.6B, suitable for text generation tasks, supporting a 32k context length and multilingual processing.

This is a 14B-parameter large mathematical language model open-sourced by NVIDIA. It has been quantized by llama.cpp and can run efficiently under different hardware conditions.

The first standalone model of the Fluently Language Model project, a 32.5B parameter causal language model supporting multiple languages and tasks.

Tülu 3 is a leading family of instruction-following models, offering fully open-source data, code, and training methodologies as a comprehensive guide to modern technology. Version 3.1 features improvements in the reinforcement learning phase, delivering enhanced overall performance.

SeaLLM-7B-v2 is a state-of-the-art multilingual large language model for Southeast Asian languages, with half the size but superior performance in multilingual tasks such as world knowledge, mathematical reasoning, and instruction following.