Easy AI Daily News Digest | December 5, 2025

Model Releases and Updates

Google releases Gemini 3 Deep Think mode

Available to Google AI Ultra subscribers, Deep Think improves complex reasoning via parallel thinking. It scores 45.1% on ARC-AGI-2, surpassing GPT-5.1's 17.6%, and supports math and science tasks.

• GoogleAI announcement
• Google DeepMind details

OpenAI launches GPT-5.1-Codex Max

Available through the Responses API and integrated into the Codex agent harness, with support for IDEs including VS Code and Cursor.
• OpenAI Devs announcement
• Cursor integration

Microsoft releases VibeVoice-Realtime-0.5B

A lightweight real-time text-to-speech model supporting English and Chinese, open-sourced on Hugging Face.
• Hugging Face model page

Nous Research releases Hermes 4.3

Built on ByteDance Seed 36B with performance close to Hermes 4 70B, trained via the Psyche network, with MoE support.
• Nous Research blog

Mistral Large 3 becomes leading open coding model

Ranked #1 on lmarena among open models, available via Ollama Cloud.
• Mistral AI announcement

Research and Technical Progress

• Titans long-context memory architecture (Google): combines RNN efficiency with Transformer performance, supports 2M+ tokens. Announcement
• TorchAO MoE quantization: introduces MoEQuantConfig for quantizing mixture-of-experts models. PyTorch PR
• VATTENTION paper: first sparse attention mechanism with (ϵ, δ) guarantees, improving long-context processing. arXiv paper
• STRAW: sample-tuned rank-augmented weights that dynamically adjust model weights, mimicking neuromodulation. Substack article
• Fast ODE solver for diffusion models: generates 4K images in 8 steps with quality comparable to 30-step DPM++2M SDE. Hugging Face Space | arXiv paper

Industry News and Funding

• Anthropic acquires Bun as Claude Code revenue reaches the $1 billion milestone. Anthropic news
• Cristiano Ronaldo invests in Perplexity, framed as "sparking curiosity worldwide." Announcement
• Harvey raises $160M Series F at an $8 billion valuation, serving 700+ law firms with legal AI. Tweet
• Antithesis raises $105M led by Jane Street, focused on deterministic simulation testing of AI-generated code. Tweet

Community Discussions

• Users on OpenAI Discord report GPT-5.1 finding code bugs that Gemini 3 missed. Discussion
• Reddit users find the Z-Image model still filters sensitive content (gore/nudity) despite claims of no censorship. Reddit thread
• Debate on AI's impact on tech jobs, with many arguing AI will change roles rather than replace them. Reddit thread
• LocalLlama users test VibeVoice-Realtime, noting some Mandarin accent issues. Reddit thread
• Heavy discussion of Gemini 3 Deep Think's ARC-AGI-2 score of 45.1% vs GPT-5.1's 17.6%. Reddit thread

Tools and Platforms

• OpenRouter releases State of AI report: analyzed 100 trillion tokens; 50% of open-model usage is roleplay, 50% of paid-model usage is coding, and Claude handles 60% of coding workloads. Report
• Windsurf integrates GPT-5.1-Codex Max with Low/Medium/High reasoning levels, free trial for paid users. Announcement
• mcp-apps-sdk open-sourced by General Intelligence Labs, enabling ChatGPT apps embedding on other platforms. GitHub repo
• tinygrad fixes train_step: PR addresses unused input tensors, improving training efficiency. PR
• DSPy community proposal to integrate Claude Code natively using its Read/Write/Terminal tools. Discussion

Benchmarks and Performance

• Gemini 3 Deep Think: 45.1% on ARC-AGI-2, 2.5x GPT-5.1's 17.6%.
• Mistral Large 3: ranked #1 in lmarena coding among open models.
• DeepSeek V3.2: Baseten serving metrics show TTFT of 0.22s and 191 tokens/sec; strong lmarena rankings in math, law, and science. Tweet
• GPT-5.1-Codex Max: positive user feedback on code quality and efficiency in Cursor and other IDEs.

---

*Source: Easy AI educational project.*