Multimodal foundation model releases accelerate
Key Questions
What new multimodal models were released recently?
Black Forest Labs launched FLUX 3 for video with native audio, while Kimi K3 (2.8T params) and Inkling (975B MoE) from Mira Murati's lab entered the market.
Why was Kimi K3's open weights release significant?
It topped benchmarks like Nemotron, was released early on Hugging Face, and enabled day-0 hosting on platforms like Together AI, though requiring substantial VRAM.
How does Opus 5 compare in pricing and capabilities?
It is offered at half Fable's price with capabilities close to Fable 5, though security concerns persist around its performance.
What benchmarks rank the latest foundation models?
A contamination-resistant coding benchmark places Fable 5 first, followed by GPT-5.5 and Claude Opus 4.8.
When are Grok updates expected?
Grok 4.5 is now widely available across platforms, with versions 4.6 and 4.7 anticipated within weeks.
What hosting considerations apply to large models like Kimi K3?
Real-world costs include 1.5TB VRAM for 8xB200 setups, with feasibility for fine-tuning and distillation being actively discussed.
How are Chinese models performing against Western leaders?
Kimi K3 outperforms some established models and offers open weights, while GLM 5.2 demonstrated strong security by repelling attacks.
What trend is accelerating in multimodal foundation models?
Rapid releases of unified multimodal systems like FLUX 3 and Inkling are expanding capabilities in video, audio, and large-scale reasoning.
Black Forest Labs launches FLUX 3. Kimi K3 (2.8T params) tops Nemotron, open weights released early on HuggingFace. Inkling (975B MoE) from Mira Murati's lab. Opus 5 at half Fable's price but security concerns. Grok 4.5 widely available. Claude Opus 5 partial downtime highlights reliability risks; another outage on July 27. Kimi K3 now available on Vercel AI Gateway and Telnyx. Analysis: Kimi K3 signals convergence toward open-weight models. Apple expands Foundation Models framework. Un-0 launches oscillator-based image generation model claiming 1000x energy reduction. Multiverse raises $570M for model compression. Moonshot AI seeks Nvidia Blackwell chips for K4 model. New: MiniMax H3 open multimodal model generating 2K video with native stereo sound, unified text/image/audio input.