Search the full model catalog.

Cohere Labs' open-weight 8B chat model, tuned for 23 languages including Arabic, Hindi, Persian and Vietnamese. The weights are under a non-commercial CC-BY-NC license.

A 12B model that Mistral AI trained jointly with NVIDIA, instruction-tuned for chat in many languages and for function calling. The card calls it a drop-in replacement for Mistral 7B.

A 1.7B-parameter English chat model from Hugging Face's SmolLM2 family, built small enough to run on-device and tuned with SFT and DPO for instruction following and function calling.

A math and science model built on Mistral 7B for problems that take several steps of reasoning. Mistral AI reports 56.6% on the MATH benchmark for it.

A 7B math model from DeepSeek, instruction-tuned on 776K English and Chinese examples. It writes step-by-step solutions and, with the recommended prompt, puts the final answer in \boxed{}.

A 3.8B-parameter Phi-3.5 chat model from Microsoft, post-trained for multi-turn and multilingual conversation across 23 listed languages. Released under the MIT license.

The 3.09B-parameter Qwen2.5 chat model that Qwen added for mobile applications. Its license, the Qwen Research License, allows non-commercial research and evaluation only.

The smallest instruction-tuned Qwen2.5 model, with 0.49B parameters and about 32,768 input tokens on Wiro. It suits routing and tagging jobs where the reply is short and the format is fixed.

At 1.54B parameters, the smallest Qwen2.5 chat model to pass 50 on both MATH and HumanEval in Qwen's tables (55.2 and 61.6). It is licensed under Apache 2.0.

Describes images and answers questions about them in plain text. It is the image checkpoint of DAMO-NLP-SG's VideoLLaMA 3, with a SigLIP-derived encoder that takes images of varying sizes.

Answers questions about a video clip or describes it in text. DAMO-NLP-SG built it on Qwen2.5-7B with a SigLIP-based vision encoder, and its training centers on image-text data.

Reads Turkish license plates in a photo and returns them as text, along with a copy of the image where each car is numbered and each plate is boxed.

Edits an existing image from a written instruction such as "add a beautiful sunset", with no mask to draw. Fine-tuned from Stable Diffusion 1.5 by Tim Brooks, Aleksander Holynski and Alexei A. Efros.

PixArt-Σ consists of pure transformer blocks for latent diffusion: It can directly generate 1024px, 2K and 4K images from text prompts within a single sampling process.

Generate synchronized audio for videos using AI. Enhance video content with matching sound effects or create custom audio from text descriptions.

UltraPixel is designed to create exceptionally high-quality, detail-rich images at various resolutions, pushing the boundaries of ultra-high-resolution image synthesis.

Qwen2's 7B chat model from June 2024, aligned with supervised fine-tuning and DPO. Qwen reports 79.9 on HumanEval and 77.2 on C-Eval; on Wiro it takes a prompt plus session history.

Writes English and Chinese math solutions either as plain derivations or with Python-assisted steps. Qwen reports 83.6 on MATH for the 7B model with chain-of-thought and 85.3 with tools.

A 1.5B-parameter Qwen2.5-Math chat model that solves English and Chinese math problems step by step and boxes the final answer. On MATH, Qwen reports 79.7 for it with tool-integrated reasoning.

A 7.61B-parameter Qwen2.5-Coder chat model for generating and fixing code. On Wiro it reads up to 32,768 input tokens per request, close to twice the limit of its 32B sibling.