Search the full model catalog.

Qwen2.5-14B-Instruct is a large language model by Alibaba’s Qwen team, optimized for instruction-following tasks with 14 billion parameters. It offers strong reasoning, multilingual capabilities, and efficient performance, making it suitable for chatbots, content creation, and various AI-driven applications.

Qwen reports 83.1 on MATH and 74.5 on Arena-Hard for this 32.5B-parameter Qwen2.5 chat model. On Wiro a request holds about 17,000 input tokens, with JSON mode and function calling.

Sana 1.6B adapted by NVIDIA for 2K text-to-image output: images around 2048 × 2048 in BF16, from prompts in English, Chinese, emoji or a mix.

NVIDIA's 1.6B-parameter Sana text-to-image model in BF16, built for 1024px images. It accepts English, Chinese and emoji prompts, and the 2K and 4K Sana checkpoints were fine-tuned from it.

Half-precision (FP16) weights for Sana 1.6B at 1024px. NVIDIA's model pairs a linear diffusion transformer with a Gemma 2 text encoder and a 32x compression autoencoder.

The 4K checkpoint of NVIDIA's 1.6B Sana model, fine-tuned from the 1024px BF16 release to generate images up to 4096 × 4096, with prompts in English, Chinese or emoji.

The largest Qwen-based R1 distill, built by DeepSeek on Qwen2.5-32B. DeepSeek's results put it ahead of OpenAI o1-mini on AIME 2024, MATH-500, GPQA Diamond and LiveCodeBench.

The mid-sized R1 distill: Qwen2.5-14B fine-tuned on DeepSeek-R1 samples. DeepSeek reports results close to or above OpenAI o1-mini on most of the benchmarks it lists.

DeepSeek's R1 distill on a Llama 3.1 8B base. It writes out a chain of thought before answering and is the only Llama option among the smaller R1 distills.

Qwen2.5-Math-7B fine-tuned by DeepSeek on samples from DeepSeek-R1. A 7B model that shows its working on math problems; DeepSeek reports 92.8 on MATH-500, above OpenAI o1-mini.

DeepSeek's smallest R1 distill: Qwen2.5-Math-1.5B fine-tuned on samples generated by DeepSeek-R1. It works through a math problem step by step before it gives the answer.

A 9B-parameter model from the EU-funded EuroLLM project, instruction-tuned for chat and translation in all 24 official EU languages plus 11 others, among them Turkish and Arabic.

utter-project/EuroLLM-1.7B-Instruct is a 1.7 billion parameter multilingual AI model designed to understand and generate text in all European Union languages and additional relevant languages. It has been trained on 4 trillion tokens from diverse sources and further instruction-tuned on EuroBlocks to enhance its capabilities in general instruction-following and machine translation.

An 8B clinical chat model from M42, fine-tuned from Llama 3 8B Instruct to answer the medical questions that general-purpose assistants are often tuned to avoid.

A 3.21B-parameter Llama 3.2 model that Meta distilled from Llama 3.1 8B and 70B, for multilingual chat and summarization. On Wiro it takes up to 128 function tools and can reply with a JSON object.

Describes a video clip in Turkish text. The model underneath, ShareGPT4Video-8B, was fine-tuned from LLaVA-Next-8B on detailed video captions written by GPT-4V.

Writes an English paragraph about what a video shows and how the camera moves. It runs ShareGPT4Video-8B, which reads 16 frames from the clip as a single 4x4 grid image.

Code Llama 34B Instruct is Meta's 34-billion-parameter code model tuned to follow chat instructions. Meta groups the 34B with the 70B as the sizes that return its best results.

The 7B instruction-tuned Code Llama from Meta writes and explains code from plain-English requests. It is the smallest Instruct size in the family.

The 8-billion-parameter instruction model of InternLM3, from Shanghai AI Laboratory. It replies directly in normal mode or works through a long chain of thought in deep thinking mode.