Which New AI Models Match GPT-5? Top Image Models, AI Quests, and Video-to-3D | AI News

TL;DR

Baidu’s Ernie X1.1 and the UAE-built K2 Think are presented as open-source models that rival leading closed systems such as GPT-5. K2 Think has 32 billion parameters and generates almost 3,000 tokens per second, while other releases handle image depth, video-to-3D point clouds, 4K images, and noisy-audio transcription. Read on to compare their capabilities, availability, and practical uses.

Transcript

AI never sleeps, and this week has been absolutely insane. We have several new open-source models that are just as good as the top closed models out there, including GPT5 and Gemini 2.5 Pro. Nvidia releases an AI that's really good at predicting the lighting of a scene, so you can add objects that blend in seamlessly. The best image generator and e... Read More

Key Insights

• Ernie X1.1 from Baidu performs on par with GPT-5 and Gemini 2.5 Pro on average benchmarks, and it is free to try at ernie.baidu.com with tools like web search, document upload, and image analysis.
• K2 Think is an open-source reasoning model developed in the UAE that matches leading models from OpenAI and DeepSeek at just 32 billion parameters, versus their hundreds of billions, and excels especially at math.
• K2 Think outputs almost 3,000 tokens per second, making it one of the fastest open-source thinking models available, and it can be downloaded on Hugging Face with separate repos for inference and fine-tuning.
• FE2E from Alibaba predicts both the normal (surface orientation) and depth of an image, outperforming other models on average for both tasks while requiring the smallest training data size.
• WinT3R (Winter) generates a 3D point cloud from a video by processing frames in chunks, creating local point maps that merge into a global map, useful for VR scenes, real estate, and interior design.
• Ernie X1.1 hallucinates far less than other leading models; given a false premise about Taylor Swift and Japan's 'Queen Elizabeth,' it correctly flags the errors instead of playing along.
• Ernie X1.1 is built on Ernie 4.5 and enhanced with extensive mid-training and post-training including reinforcement learning, a pipeline similar to how DeepSeek R1 improved on DeepSeek V3.
• This week's releases also include Seedream 4 with 4K image generation, Tencent's HunyuanImage 2.1 for offline use, an anime background remover with 99.5% pixel accuracy, Qwen3 ASR transcription for noisy audio, and Stability's Stable Audio 2.5 music generator.

Questions & Answers

Q: What new open-source AI models match GPT-5?

Baidu’s Ernie X1.1 performs on par with GPT-5 and Gemini 2.5 Pro on average benchmarks. K2 Think also delivers frontier-level reasoning at 32 billion parameters, matching leading reasoning models from OpenAI and DeepSeek.

Q: What is K2 Think, and how well does it perform?

K2 Think is an open-source reasoning model developed in the UAE with 32 billion parameters. It matches much larger leading reasoning models and performs especially well at math, although it is weaker at coding and only on par for science.

Q: How fast is K2 Think?

K2 Think produces almost 3,000 tokens per second, making it one of the fastest open-source thinking models described on the page. It answered an Olympiad-level math problem in about a second while still performing reasoning before responding.

Q: What does Alibaba’s FE2E model do?

FE2E predicts both image depth and surface normals, which represent the orientation of surfaces in a scene. It performs accurately on complex and cluttered images, including background details such as stairs and gates that other depth estimators may miss.

Q: How does FE2E compare with other depth and normal estimation models?

FE2E outperforms the compared models on average for both depth estimation and normal estimation. It also required the smallest training-data size among the models shown in the comparison charts.

Q: How does WinT3R turn a video into a 3D scene?

WinT3R, also called Winter, processes a video as a stream of image frames and generates local point maps chunk by chunk. It then merges those local maps into a global 3D point cloud that can reconstruct a room and its objects.

Q: What can WinT3R’s 3D point clouds be used for?

WinT3R can reconstruct detailed scenes from a video or a few input photos, producing a point cloud rather than a 3D mesh. The transcript suggests applications such as VR scenes, real estate, and interior design.

Q: What new image, audio, and transcription tools are covered?

The releases include Seedream 4 for 4K image generation, Tencent’s HunyuanImage 2.1 for offline use, and an anime background remover with 99.5% pixel accuracy. They also include Qwen3 ASR for noisy or low-quality audio with language autodetection and Stable Audio 2.5 for music generation.

Summary & Key Takeaways

• Multiple new open-source AI models released this week rival the best closed models. Alibaba's FE2E predicts image depth and surface normals more accurately than competitors while using the least training data, and WinT3R reconstructs detailed 3D point clouds from ordinary video frames by merging local maps chunk by chunk.

• K2 Think, built in the UAE, delivers frontier-level reasoning at only 32 billion parameters, matching much larger OpenAI and DeepSeek models and excelling at math. It runs extremely fast at nearly 3,000 tokens per second, is free to try online, and can be downloaded from Hugging Face with inference and fine-tuning repos.

• Baidu's Ernie X1.1 performs on par with GPT-5 and Gemini 2.5 Pro, offering strong factual accuracy, instruction following, and agentic ability while hallucinating notably less. Additional releases include Seedream 4 with 4K images, HunyuanImage 2.1, an anime background remover, Qwen3 ASR, and Stable Audio 2.5.