# Code Generation Optimization

AReaL is an asynchronous reinforcement learning training system developed by Ant Group, designed specifically for large inference models, supporting fast training and cutting-edge performance.

A math and code reasoning model trained through reinforcement learning, based on DeepSeek-R1-Distilled-Qwen-7B, excelling in mathematical and code reasoning tasks

SWE-agent-LM-32B is an open-source software engineering language model, fine-tuned based on Qwen/Qwen2.5-Coder-32B-Instruct, and is specifically designed for software engineering tasks.

Falcon-H1 is a high-performance hybrid architecture language model developed by TII, combining the strengths of Transformers and Mamba architectures, supporting English and multilingual tasks.

Falcon-H1 is a decoder-only causal model with a hybrid Transformers + Mamba architecture developed by TII, supporting English and multilingual tasks.

GLM-4-9B-0414 is a lightweight member of the GLM family with 9 billion parameters, excelling in mathematical reasoning and general tasks, providing an efficient solution for resource-constrained scenarios.

Granite-4-Tiny-Preview is a fine-grained Mixture of Experts (MoE) instruction-tuned model with 7 billion parameters, developed based on Granite-4.0-Tiny-Base-Preview, suitable for general instruction-following tasks.

A 14B-parameter code generation and comprehension model based on the Qwen2.5 architecture, integrated through the Model Stock method by combining multiple specialized coding models

A hybrid model based on Qwen2.5-Coder-14B-Instruct, merged using the mergekit tool with multiple specialized coding models

A quantized version based on burtenshaw/GemmaCoder3-12B, optimized for code generation tasks, supporting multiple quantization levels to adapt to different hardware requirements.

Z1 is a large language model based on Qwen2.5-Coder-7B-Instruct, focusing on efficient reasoning through thought migration.

Qwen2.5-Coder-32B-instruct is an instruction fine-tuned version based on Qwen2.5-Coder-32B, specializing in code generation and understanding tasks.

Combines the top 32B reasoning and coding models from the open-source community, using the SCE merging technique with Qwen2.5-Coder-32B as the base model for fusion.

A 32B-parameter large language model merged from the Qwen series, specializing in code generation and comprehension tasks

This model is a fine-tuned version of google/gemma-3-4b-it on the codeforces-cots dataset, primarily used for code-related tasks.

OpenCerebrum-1.0-7b-SFT is a statically quantized model based on Locutusque/OpenCerebrum-1.0-7b-SFT, supporting multiple quantization versions, suitable for text generation and Q&A tasks in fields such as code generation, mathematics, chemistry, and biology.

An optimized version based on Qwen2.5-Coder-7B-Instruct, focused on eliminating refusal behaviors

Llamacpp imatrix quantized version based on the Qwen2.5-Coder-1.5B-Instruct-abliterated model, suitable for code generation and text generation tasks.

Granite-3.0-8B-Instruct is an 8-billion-parameter model fine-tuned from Granite-3.0-8B-Base, trained using a combination of various open-source instruction datasets and internally synthesized datasets.

PowerLM-3B is a small language model with 3 billion parameters, trained using the Power learning rate scheduler, and demonstrates excellent performance across multiple benchmarks including natural language multiple-choice, code generation, and mathematical reasoning.

A large language model fine-tuned based on the CodeLlama-13b-hf architecture, specifically optimized for API call generation tasks

A code generation model fine-tuned on 100,000 high-quality ranked data points using reference-free monolithic preference optimization, based on the Qwen/CodeQwen1.5-7B model

This is a merged model focused on code generation and application development, created by fusing multiple high-quality models

OpenCodeInterpreter is an open-source code generation system that enhances code generation capabilities by integrating code execution and iterative optimization functions

DeepMagic-Coder-7b is a merged version of the DeepSeek-Coder and Magicoder models, focusing on code generation and programming tasks.

TinyMistral-6x248M is a Mixture of Experts system that integrates 6 TinyMistral variants using the LazyMergekit tool, pre-trained on the nampdn-ai/mini-peS2o dataset

The EveryoneLLM series models are created by the community and serve the community. This is a specialized coding model fine-tuned based on deepseek-coder-33b-base.

phixtral-2x2_8 is the first Mixture of Experts (MoE) model built upon two microsoft/phi-2 models, outperforming each individual expert model.

Instruction-tuned model based on Mistral-7B, focusing on following complex instructions rather than daily chat, using llama-2 chat format

A version fine-tuned by Open-Assistant based on Meta's CodeLlama 13B large language model, supporting English, with a new RoPE Theta value (1e6 instead of 1e4).

A large language model fine-tuned based on CodeLlama-34B-Python, achieving 69.5% pass@1 on HumanEval, surpassing GPT-4's performance

A 125M-parameter code generation model based on the GPT-Neo architecture, fine-tuned on the CodeClippy dataset, supporting sequence generation up to 2048 tokens