AI News: 6 New Models & A Nano Banana Killer

Summary

The video reviews a busy week in AI, covering image tools, large language models, open-weight releases, and product updates.

Image generation and editing

• Krea 3 was the presenter’s standout image tool, particularly for editing existing images. Users can select specific objects or areas and change them—such as a phone’s background, text, or a hat logo—while leaving the rest of the image largely intact. The presenter found the workflow fast and the edits cleaner than comparable tools, including ChatGPT’s image feature and Nano Banana. Krea 3 was available in public beta.
• FLUX 3 Image Klein is an open-weight image model that can generate and edit images locally. The video notes that it can run on consumer hardware, such as an RTX 4070. A Hugging Face demo offers free access, though it may be slower.
• Qwen Image 2.1 Turbo is a faster version of Alibaba’s image model, trading some quality for speed. It is also available to try on Hugging Face. Overall, the presenter was more impressed by the editing capabilities of these models than by their image generation.

Design tool

GenSpark Design can turn reference screenshots into a multi-screen mobile app design, including feeds, tool pages, search, filters, and news sections. Generated components remain layered and editable. The workspace also supports creating other assets and collaborating with a team. Designs can be exported or passed to GenSpark Code to begin implementation.

This segment was sponsored by GenSpark.

Major language models

• Gemini 4 Argon was described as competitive with leading models on many benchmarks while costing less to use than some rivals. The presenter tested it on a coding and game-generation task, as well as creating a 3D-printable bust from a photo. He found the bust convincing, while the generated game was functional but not especially polished visually. A separate video provides a fuller review.
• Meta Muse Wonder reportedly scored strongly on benchmarks, including 79% on a coding benchmark. The presenter’s own tests were more mixed: its generated game included the requested mechanics but had weaker visuals, and it ranked 14th on his image-generation leaderboard. He noted that its price was in a similar range to Claude Opus.
• Muse Spark Open is an open-weight release described as an 816-billion-parameter model with a one-million-token context window. Meta and Apple reportedly optimized it to run on an M5 Ultra Mac Studio, though the presenter noted that such a system is expensive. Its benchmark performance was strongest compared with other open-weight models, rather than the newest closed models.
• Microsoft MAI-Base-2 was announced as an 80-billion-active-parameter model, but was not yet available for independent testing.
• NVIDIA Nemotron 3.5 Super VL is a 120-billion-parameter open-weight model. Its main reported advantage is speed; the presenter said its benchmark results did not match those of some leading open models.
• Ultramarine 1 Flash was presented by its developer as a low-cost model with strong knowledge-work performance. The presenter could not test it because it was not yet available through OpenRouter.

OpenAI updates and other models

The presenter found the latest part of OpenAI’s “28 days of releases” underwhelming. Updates included:

• Predicting a user’s next prompt.
• Creating and delegating work to a “dot” from a phone.
• Increasing GPT 6.1 Soul’s default context length in Codex and Work.
• Reducing desktop-app disk usage.
• Adding computing capacity to reduce overload errors.

He characterized several of these as minor changes or reliability fixes rather than major new features.

The video also covered two new decision-making models: Microsoft Decision One and Bespoke Labs’ Nimble V4 and V4 Max. These models assign probabilities or scores to fixed options, such as yes/no choices, multiple-choice answers, AI responses, or agent actions. The presenter described them as similar in purpose, with accuracy as the main differentiator.

Voice, agents, and broader AI news

• Grok Voice 3.0 offers a sandbox for building voice agents. The presenter demonstrated a Future Tools support bot that used the site’s information to answer questions, highlighting its quick responses and natural-sounding voices.
• ElevenLabs Instant Soundtracks can analyze an uploaded video and generate a soundtrack with synchronized sound effects.
• Facebook added AI-based, real-time personalization of users’ feeds. Google introduced a Circle to Search feature intended to help identify suspicious text messages.
• Kylon is a workspace where multiple AI agents and people can coordinate through a shared messaging system.
• Microsoft’s Xbox Labs is exploring AI-assisted game mashups and game generation. Epic Games is also shifting toward broader AI adoption. The presenter said AI-assisted coding in games may become common, while AI-generated creative assets may face more resistance.
• Apple showed new devices, including an Apple TV 4K and a screen-equipped smart-home device referred to in the subtitles as “HomePod View.” The presenter had not tested them.

Reviews, guides, and demonstrations

The video included:

• A step-by-step Krea 3 editing demo, showing how to select and change objects, text, and logos in an image.
• Demonstrations of how to try FLUX 3 Image Klein and Qwen Image 2.1 Turbo through Hugging Face.
• A GenSpark Design demo showing how website screenshots can be converted into an editable mobile app design.
• Tests of Gemini 4 Argon, Muse Wonder, Muse Spark Open, and Nemotron using image-generation and coding/game tasks, plus a 3D-printing task for Gemini.
• A Grok Voice 3.0 demo showing how to create and use a simple website-support agent.
• A reference to a separate, more detailed breakdown of Gemini 4 Argon.

Main speaker: Matt Wolfe

Sources and organizations discussed: Krea, Black Forest Labs, Alibaba Qwen, GenSpark, Google, Meta, Microsoft, NVIDIA, Paper Instruments, OpenAI, Bespoke Labs, xAI, ElevenLabs, Facebook, Kylon, Xbox/Microsoft, Epic Games, and Apple.

Benchmark references included Artificial Analysis and coding-benchmark results cited in the video. Theo was also quoted on his model rankings.