Welcome back!

The open-weight Kimi K2.5 from Moonshot AI is shaking up the leaderboards. It’s a trillion-parameter model orchestrating hundreds of agents in parallel, outpacing even Gemini 3 Pro. Its architecture hints at a fresh scaling method driven by agent collaboration. Let’s unpack…

Today’s Summary:

• 🔥 Kimi K2.5 redefines open-weight scaling

• ⚡ FLUX.2 Klein: sub-second image generation

• 💬 OpenClaw open-source AI assistant

• ✍️ OpenAI launches Prism workspace

• 📜 Claude introduces a new constitution

• 📢 OpenAI prepares ChatGPT ads

• 🛠️ 2 new tools

TOP STORY

Open-weight Kimi K2.5 tops Gemini 3 Pro with parallel agent system

The Summary: Moonshot AI has released Kimi K2.5, an open-weight multimodal model with 1 trillion parameters that coordinates up to 100 AI agents working simultaneously on complex tasks. It uses Parallel-Agent Reinforcement Learning to automatically divide work among specialized sub-agents, reducing execution time by up to 4.5x.

Key details:

• Ranks #1 open model on Vision Arena, tied for #3 overall

• Can execute up to 1,500 coordinated tool calls across 100 sub-agents

• Trained on 15 trillion mixed visual and text tokens, achieving 76.8% on SWE-Bench Verified (vs Claude 4.5 Opus at 80.9%)

• The 595GB model file requires substantial hardware, prompting developers to request smaller versions around 70B parameters for local deployment

• Developers praise it as the first open model truly competing with top labs

Why it matters: The multi-agent architecture represents a new scaling method using agent coordination, which may offer new capability gains as traditional approaches face diminishing returns. The tension between open weights and deployment reality shows that most developers lack sufficiently powerful hardware to run these models locally, making "open" still more theoretical than practical for the community.

FROM OUR PARTNERS

Talk Your Prompts Into Shape

Speak fuller prompts. Get better answers.

Stop losing nuance when you type prompts. Wispr Flow captures your spoken reasoning, removes filler, and formats it into a clear prompt that keeps examples, constraints, and tone intact. Drop that prompt into your AI tool and get fewer follow-up prompts and cleaner results. Works across your apps on Mac, Windows, and iPhone. Try Wispr Flow for AI to upgrade your inputs and save time.

Try Wispr Flow

BLACK FOREST LABS

FLUX.2 Klein generates images in under one second

The Summary: Black Forest Labs released FLUX.2 [Klein], an open-source AI image generator that creates visuals in under a second on local hardware. The startup, founded by former Stability AI engineers, offers two variants: a 4B parameter model under Apache 2.0 license (fully commercial) and a 9B parameter version for non-commercial use. The models run on consumer GPUs.

Key details:

• 4B variant generates in 1.2 seconds on RTX 5090, fits in 8.4GB VRAM

• Apache 2.0 license on 4B model allows free commercial use

• Unified architecture handles text-to-image, single-reference editing, and multi-reference composition natively without model swapping

• Accepts hex-code color control for precise color rendering and JSON-like structured inputs for programmatic generation

Why it matters: The distillation approach achieving four-step generation represents a substantial architecture improvement over legacy models. It is now possible to deploy production-quality image generation in local networks without compromising on speed, legal clarity, or costly cloud budgets.

Try the demo

FROM OUR PARTNERS

Work Smarter With These Prompts

Want to get the most out of ChatGPT?

ChatGPT is a superpower if you know how to use it correctly.

Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.

Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.

Download the free guide

OPEN SOURCE

Open-source AI assistant OpenClaw goes viral despite security warnings

The Summary: OpenClaw is an open-source AI assistant that operates 24/7 from messaging apps like Telegram and WhatsApp. The tool manages calendars, emails, and tasks, running locally on user hardware while calling cloud models like Claude Opus via API. Security researchers have raised concerns about the system's broad access permissions.

Key details:

• Accumulated 119K+ GitHub stars in two months

• Security researchers flagged exposed instances online, some revealing access to messages, credentials, and API keys

• Applications include automated apartment rental screening via Facebook Messenger, negotiating car purchases over email, or self-created integrations like Whisper transcription

• Some users reported $300+ API costs in two days, with the system consuming many tokens per interaction

Why it matters: OpenClaw demonstrates strong demand for useful autonomous agents. The security concerns highlight a design tradeoff: the system grants broad access to files, credentials, and commands to maximize utility. But this creates risks that matter especially for non-technical users who may not fully understand the permission scope. This gap between consumer appeal and security remains a challenge.

QUICK NEWS

Quick news

• OpenAI launches Prism LaTeX workspace for scientific writing

• Claude has a new constitution

• OpenAI to introduce ads in ChatGPT as part of new monetization plan

TOOLS

🥇 New tools

• Kilo Code Reviewer - Open source code review agent

• ChartGen - Create professional charts with insights