# ThursdAI - Dec 4, 2025 - DeepSeek V3.2, Mistral 3 Apache 2.0, OpenAI Code Red & US-Trained MOEs! Source: https://www.youtube.com/watch?v=nnGXN1JMNYk Recap page: https://rapidrecap.app/video/nnGXN1JMNYk Generated: 2025-12-05T21:09:49.233+00:00 --- ## Quick Overview The Thursday AI news roundup highlighted major open-source releases including DeepSeek V3.2 Special rivaling frontier models, Mistral 3 Apache 2.0 models with mixed reception on reasoning capabilities, and RC AI's US-trained Trinity MoE models, alongside OpenAI's reported 'Code Red' in response to competitive pressures from Gemini and other major lab developments. **Key Points:** - DeepSeek V3.2 Special achieved a gold medal on the Olympiad with 685 billion parameters, ranking as the second most intelligent open-weights model, even surpassing Claude 4.5 on some stats, and costs only 28 cents per million tokens on OpenRouter. - Mistral released Mistral 3, including Mistral Large (675B parameters, 256k context window, 41B active parameters) and smaller multimodal models (3B, 8B, 14B), all under the Apache 2.0 license. - The Mistral Large model is noted as a non-reasoning instruction model, leading to lower scores on reasoning evaluations like the Artificial Analysis intelligence score compared to reasoning models. - RC AI released Trinity, a family of fully US-trained Mixture of Experts (MoE) models under the Apache 2.0 license, including Trinity Mini (26B) and Trinity Nano (6B preview), with Trinity Large (420B parameters, 13 experts) targeting mid-January 2026 release. - OpenAI declared a 'Code Red' internally following a reported 6% daily active user drop after Gemini 3's launch, pausing side projects to focus on speed and personalization. - Whisper Thunder, revealed to be Runway's Gen 4.5, supposedly beats V3, Sora 2 Pro, and Cling on ELO scores with superior physics but lacks native audio. - Weights & Biases launched a preview LLM evaluation service allowing users to evaluate any OpenAI-compatible API hosted model directly using standard evaluation sets. **Context:** The hosts Alex Volov, Wolf, Niston, and Yam Pelleg discussed the top AI releases for the week of December 4th, 2025, covering significant advancements in both open-source and closed-source large language models, as well as breaking news concerning major lab strategies and new multimodal video generation capabilities. ## Detailed Analysis The week saw major open-source activity, led by DeepSeek's V3.2 and V3.2 Special (685B parameters, MIT license), which achieved top-tier reasoning scores, rivaling Gemini 3 Pro and costing extremely little for inference. Mistral also made headlines by releasing the Mistral 3 series under Apache 2.0, including Mistral Large (675B, MoE), which excels in coding benchmarks but is characterized as a non-reasoning model, placing it lower on pure intelligence rankings compared to models like Gemini 3. RC AI introduced Trinity, a US-trained MoE family, specifically addressing enterprise compliance needs for domestically trained models, with the large frontier model anticipated in mid-January 2026. In closed labs, OpenAI reportedly entered a 'Code Red' due to user churn following Gemini 3's release, allegedly pausing projects while working on a secret model named 'Garlic'. Amazon announced Nova 2 models emphasizing enterprise pre-training checkpoints and featuring Nova 2 Omni, a multimodal model supporting text, image, video, and speech input/output with a 1 million context window. Video generation saw Runway's Whisper Thunder (Gen 4.5) claiming superiority over Sora 2 Pro, while Cling released Video 2.6 with native audio generation. Finally, the hosts detailed a new Weights & Biases feature allowing direct evaluation of any OpenAI-compatible API model. ### Open Source Model Releases - DeepSeek V3.2 Special achieves 96% on AIM and beats Claude 4.5 on some stats, ranking as the #2 open model - Mistral 3 Large (675B) is Apache 2.0 licensed but is a non-reasoner, scoring lower on intelligence benchmarks - RC AI released US-trained Trinity Mini (26B) and Nano (6B) MoEs, with Trinity Large targeting mid-January 2026. ### Closed Lab & Enterprise News - OpenAI declared a 'Code Red' due to a 6% daily active user drop post-Gemini 3 launch, pausing side projects - Amazon launched Nova 2 series, focusing on enterprise pre-training checkpoints, and Nova 2 Omni (multimodal, 1M context window) - Anthropic acquired Bun, and OpenAI acquired Neptune AI, leading to Neptune shutting down services. ### Multimodal & Video Generation Updates - Whisper Thunder (Runway Gen 4.5) supposedly beats Sora 2 Pro on ELO scores but lacks native audio - Cling Video 2.6 features native audio generation, 1080p up to 10 seconds, and is cheaper than previous versions - Biden released updates to Cdream with multi-references and multilingual text. ### Tools & Ecosystem Updates - Weights & Biases launched LLM evaluation jobs to test mid-training checkpoints or any OpenAI-compatible API model - Wolf highlighted that competitors like HumanLoop and Neptune AI are being acquired and shutting down, contrasting with W&B's commitment to continued service. ### Interview with RC AI CTO Lucas Atkins - Trinity models address enterprise compliance needs for US-trained models, as growth is slowed by legal teams scrutinizing model origins - Trinity Large (420B parameters, 13 experts) trains now for a mid-January release - MoE models are highly inference efficient, which is crucial for RL paradigms involving massive rollouts.