Hypnos i1-8B is a professional reasoning model based on Nous Hermes 3 (Llama 3.1 8B), focusing on complex logic, chain-of-thought reasoning, and mathematical problem-solving. This model has conducted unique exploration in the field of hybrid quantum-classical machine learning, and improved the model's creativity and reasoning ability by introducing quantum noise injection.

Natural Language ProcessingTransformersEnglish

VibeThinker-1.5B is a 1.5-billion-parameter dense language model launched by Weibo AI. It is fine-tuned based on Qwen2.5-Math-1.5B and is specifically designed for mathematical and algorithmic coding problems. Trained using the 'Spectrum to Signal Principle' framework, it outperforms larger models in multiple math competition tests. The training cost is approximately $7,800, and it supports an output of up to about 40k tokens.

Natural Language ProcessingTransformersMultiple Languages

DeepMath is a mathematical reasoning model with 4 billion parameters, built on Qwen3 - 4B Thinking, combining a fine - tuned large language model and a sandboxed Python executor. It can generate concise Python code snippets for calculation steps, significantly reducing errors and shortening the output length.

Natural Language ProcessingTransformersEnglish

The 4-bit quantized version of VibeThinker-1.5B, optimized for Apple chips based on the MLX framework, is a dense language model with 1.5 billion parameters, specifically designed for mathematical reasoning and algorithm coding problems.

Natural Language ProcessingMlxMultiple Languages

Olmo 3 7B RL-Zero Mix is a language model with 7B parameters developed by Allen AI, belonging to the Olmo 3 series. This model was pre-trained on the Dolma 3 dataset, post-trained on the Dolci dataset, and optimized for mathematical, coding, and reasoning abilities through reinforcement learning.

Natural Language ProcessingTransformersEnglish

Olmo 3 7B RL-Zero Math is a 7-billion-parameter language model developed by Allen AI, specifically optimized for mathematical reasoning tasks. It is trained on mathematical datasets using the RL-Zero reinforcement learning method, which can effectively improve mathematical reasoning ability.

Natural Language ProcessingTransformersEnglish

Rnj-1 is a series of 8 billion parameter, open-weight dense models trained from scratch by Essential AI. This model is optimized for the code and STEM fields, excelling in programming, mathematical reasoning, agent tasks, and tool invocation, with capabilities comparable to the most advanced open-weight models.

Natural Language ProcessingTransformers

Wraith-8B is the first model in the VANTA Research entity series, fine-tuned based on Meta's Llama 3.1 8B Instruct. This model performs excellently in mathematical reasoning, achieving an accuracy rate of 70% in the GSM8K benchmark test. It also has a unique cosmic intelligence perspective, providing strong support for various application scenarios.

Natural Language ProcessingTransformersEnglish

metatune-gpt20b is a large language model prototype with self-improvement ability. It can generate new data for itself, evaluate its own performance, and adjust hyperparameters according to improvement indicators. This model performs excellently in scientific and mathematical understanding at the postdoctoral level and can also be used for coding tasks.

Natural Language ProcessingTransformersEnglish

A-vibe is a Russian-optimized large language model developed by Avito Tech based on Qwen3-8B-Base. Through word segmenter optimization, model slimming, and multi-stage training, it performs excellently in Russian benchmark tests and supports function calls and mathematical problem-solving.

Natural Language ProcessingSafetensorsMultiple Languages

Qwen-1.5B_THIP is a mathematical reasoning model fine-tuned using the GRPO method on the DeepMath-103k mathematical dataset based on DeepSeek-R1-Distill-Qwen-1.5B with the TRL framework. This model is specifically optimized for solving mathematical problems and has strong mathematical reasoning abilities.

Natural Language ProcessingTransformers

The FP8 quantized version of olmOCR-2-7B-1025, fine-tuned based on Qwen2.5-VL-7B-Instruct, is a vision-language model specifically designed for handling complex OCR scenarios such as mathematical equations and tables.

MultimodalTransformersEnglish

This project distills the inference ability of GPT into the Llama-3.1-8B model through an innovative two-stage training process. First, knowledge distillation and format alignment are performed through supervised fine-tuning, and then reinforcement learning is used to encourage the model to independently explore and optimize inference strategies, focusing on breakthroughs in the field of mathematical reasoning.

Natural Language ProcessingTransformersMultiple Languages

Frugal-Math-4B is a 4B parameter language model optimized for mathematical reasoning. It is trained using the Reinforcement Learning with Verification Reward (RLVR) method and can generate concise and verifiable mathematical solutions while maintaining high accuracy, significantly reducing the verbosity of reasoning.

Natural Language ProcessingTransformersMultiple Languages

QuestA is an innovative framework that enhances the reasoning ability of large language models through a question-enhanced approach. It integrates partial solutions during the reinforcement learning training process, significantly improving the model's performance on complex tasks such as mathematical reasoning, especially achieving optimal results on small-parameter models.

Natural Language ProcessingTransformers

This is a speculative model based on the EAGLE-3 speculative decoding algorithm, specifically designed to work with the Qwen/Qwen3-32B model. It improves text generation efficiency through speculative decoding technology and performs well in mathematical reasoning and general Q&A.

Natural Language ProcessingTransformers

The palmyra-mini series of models demonstrates excellent capabilities in the fields of complex reasoning and mathematical problem-solving, especially performing outstandingly in benchmark tests that require in-depth understanding and multi-step thinking. This model is proficient in solving elementary school-level math problems, achieving a high score of 0.818 in the gsm8k benchmark test, also obtaining a score of 0.818 in the MATH500 benchmark test, and a stable score of 0.6 in the AMC23 benchmark test.

Natural Language ProcessingTransformersMultiple Languages

LLaDA-MoE is a new type of mixture-of-experts language model constructed based on the diffusion principle. It is the first open-source MoE diffusion large language model, pre-trained from scratch on approximately 20 trillion tokens. It has a total of 7 billion parameters, and only 1.4 billion parameters are activated during inference. It performs excellently in tasks such as code generation and mathematical reasoning.

Natural Language ProcessingTransformers

MobileLLM-R1 is an efficient inference model series released by Meta, including three scales of 140M, 360M, and 950M. This model is specifically optimized for mathematical, programming, and scientific problems, achieving comparable or even better performance than large-scale models with a smaller parameter scale.

Natural Language ProcessingTransformersEnglish

MobileLLM-R1 is a series of efficient inference language models released by Meta, focusing on solving mathematical, programming, and scientific problems. This model can still achieve excellent performance with a relatively small parameter scale, and provides a complete training recipe and data sources to support reproducible research.

Natural Language ProcessingTransformersEnglish