**Claude 4.5 Opus Hits 4-Hour Benchmark; Rakuten’s 700B Model Challenges U.S. AI Dominance**

AI Model Releases & Benchmarks

Google DeepMind’s "Flash" Outperforms "Pro" in Agentic Reinforcement Learning (RL)

• A research scientist at Google DeepMind revealed that the "Flash" model surpasses "Pro" due to advancements in agentic RL, with benchmark results (SWE-Bench Verified) showing it outperforming GPT-5.2 and Opus 4.5.
  • Deleted post from a research scientist @ GoogleDeepMind

Claude 4.5 Opus Achieves 4-Hour 49-Minute Time Horizon on METR Benchmark

• Claude Opus 4.5 demonstrated a 50% task completion time horizon of 4 hours and 49 minutes for software engineering tasks, marking significant progress in long-horizon AI capabilities.
  • Claude 4.5 opus achieves metr time horizon of 4 hours 49 mins

Mistral’s Vibe (with devstral 2) vs. Claude Code on SWE-bench-mini

• Mistral’s Vibe (with devstral 2) scored 37.6% on SWE-bench-mini, narrowly trailing Claude Code’s 39.8%, with results statistically indistinguishable.
  • Mistral's Vibe (with devstral 2) vs Claude Code on SWE-bench-mini: 37.6% vs 39.8%
  • External comparison thread

New AI Tools & Frameworks

NVIDIA Releases NitroGen: Open-Source Vision-to-Action Model

• NitroGen is a new model that plays video games from raw frames using large-scale imitation learning on human gameplay data, showcasing potential for embodied AI.
  • NitroGen: NVIDIA's new image-to-action model
    • Hugging Face Model
    • Project Website
    • Paper
  • Key Highlights of NVIDIA’s New Open-Source Vision-to-Action Model: NitroGen
    • Hugging Face Model

Aider-ce: Enhanced AI Coding CLI with Agent Mode & Multi-Codebase Processing

• Aider-ce, the successor to Aider, introduces agent mode, MCP (Multi-Codebase Processing), and skills for cost-efficient, token-light coding assistance.
  • Aider-ce is the new Aider (easiest way to learn how a barebones AI coding CLI works)
    • GitHub Repository

FlashHead: 50% Faster Token Generation for Small Language Models (SLMs)

• FlashHead is a drop-in replacement for LM heads, using information retrieval to accelerate token generation without sacrificing accuracy.
  • FlashHead: Up to 50% faster token generation on top of other techniques like quantization
    • Hugging Face Demo
    • GitHub

Product Updates & Integrations

Roo Code 3.36.7–3.36.16: Gemini 3 Flash Preview & Native Tools by Default

• The update introduces native tools for more providers/models, adds Gemini 3 Flash preview, and improves chat error troubleshooting.
  • Roo Code 3.36.7-3.36.16 Release Updates
    • Release Notes

Qwen-Image-Layered: Photoshop-Grade Layered Image Editing

• Qwen’s new model enables RGBA layer manipulation with prompt-controlled structure and infinite decomposition for advanced image editing.
  • Qwen released Qwen-Image-Layered on Hugging Face
    • Hugging Face Model

Industry & Strategic Developments

Google Faces Compute Constraints, Forms Allocation Council

• Google’s compute crunch has led to a high-level council managing capacity, impacting AI strategy for the next 12–18 months despite heavy investments.
  • Even Google is compute constrained and that matters for the AI race
    • The Information Article

Rakuten to Release 700B Open-Weight Model in Spring 2026

• Japan’s Rakuten plans a 700B-parameter open-weight model, potentially rivaling Chinese models and pressuring U.S. firms to scale up.
  • Japan's Rakuten is going to release a 700B open weight model in Spring 2026
    • Yahoo Japan News

Emerging Models & Previews

MiniMax-M2.1: Next-Gen Model with 3D Particle System Demo

• M2.1, the upcoming version of MiniMax-M2, was previewed with an interactive 3D particle system, hinting at improved performance and efficiency.
  • Just pushed M2.1 through a 3D particle system. Insane！