AI News: Vibe Jam, Mistral Small 3.1 and New Audio Models

In this March 21, 2025 AI news roundup, I start with vibe coding. After Levels.io demonstrated an AI-assisted flight simulator, he announced Vibe Jam, a competition for web-based multiplayer games made with this approach. The submissions show how quickly people are experimenting with playable projects.

One multiplayer entry combines a Fortnite-style shooter with Minecraft-like visuals and is already being play-tested. Other examples include a safari driving game, a line-rider-style game, a puzzle reportedly built in hours, Vibe Tanks, a food fight simulator, and an air-traffic-control game that evolves from a simple 2D view into a fuller 3D map.

Mistral released Mistral Small 3.1, a 24-billion-parameter multimodal open-weight model. The company's displayed GPQA Diamond chart puts it ahead of several named models on that test while showing low per-token latency. This is a vendor comparison on one benchmark, not evidence that it leads every task.

Mistral says the model can run on a single RTX 4090 or a Mac with 32 GB of RAM, making local inference interesting, and lists a 128,000-token context window. I see it as a foundation that could be trained further for reasoning; actual performance and hardware requirements depend on the setup.

Claude gained web search. I see that as useful for coding in particular: the assistant can look for current API documentation, library information and recent bugs rather than relying only on what it already knows. My view that Claude was a leading coding model is an opinion from the time of this video.

OpenAI announced two speech-to-text models, GPT-4o Transcribe and GPT-4o mini Transcribe, alongside a new text-to-speech model. The company says the transcription models improve on Whisper in its evaluations; benchmark results and language coverage should be read in the context of those tests.

The text-to-speech demo at OpenAI.fm lets users specify not only the words but how they should be spoken, including voice, tone, pacing and emotion. I play a dramatic example and note that the interface can also display code for incorporating the feature into an application.

OpenAI and Teenage Engineering also promoted a creative text-to-speech competition, with OB-4 devices as prizes. This was an announcement from the March 2025 news cycle, so the contest details should be understood in that historical context.

Windsurf's Wave 5 focuses on Windsurf Tab. In a company clip, its representative says the feature combines autocomplete, multi-line edits, navigation and imports in one tool. The announcement says free users also receive unlimited tab completion; those are the terms presented at the time.

Krea AI introduced video training for Wan 2.1. The idea is to train on your own videos so generated clips can reflect particular styles, objects or movements. Separately, NotebookLM added interactive mind maps derived from supplied documents, offering another way to explore the material alongside audio overviews.

Tencent's Hunyuan team announced Hunyuan3D 2.0 MV for multi-view generation and Hunyuan3D 2.0 Mini. I point to their potential for creating 3D characters for games, motion graphics and videos; the models shown were made available for download.

Stability AI introduced Stable Virtual Camera, which generates a video with camera movement from 2D image inputs. The demo shows moves such as zooming and changing viewpoint. I see it as a step toward more accessible AI filmmaking, though a demonstration is not a finished film production system; the stated release was for non-commercial use.

Gemini added Canvas for writing, editing and running code in the browser. I compare the experience to similar capabilities already available in Claude and ChatGPT, and suggest using it to iterate on HTML or JavaScript projects.

LG AI Research announced EXAONE Deep, a reasoning model family in 32B, 7.8B and 2.4B sizes. LG's announcement claims the 32B model led its AIME comparison while using a fraction of a competitor's parameters; the smaller models are more suitable for local use. These are historical vendor-reported benchmark claims, not a general ranking of model capability.