Anthropic Unveils "Claude for Small Business" for SMBs

On May 13, Anthropic announced "**Claude for Small Business**," a package aimed at small and medium-sized businesses. It integrates Claude with tools that small businesses use every day — QuickBooks, PayPal, HubSpot, Canva, DocuSign, Google Workspace, and Microsoft 365 — enabling automation of tasks such as payroll planning, monthly closing, sales campaigns, and invoice follow-ups.

It can be enabled with a single toggle on Claude Cowork and connects with existing tools. The offering ships with 15 prebuilt agentic workflows, and users only need to handle approvals. Anthropic co-founder and President Daniela Amodei positioned AI as "the first technology that can close the gap between big and small businesses."

TechCrunch analyzes the move as "evidence that the AI platform wars are expanding downmarket," noting that the next major battleground is not Fortune 500 companies but the 36 million small businesses in the United States.

Fractile Raises $220 Million to Speed Up AI Inference

AI chip startup **Fractile** has raised $220 million (approximately ¥33 billion), the Wall Street Journal reports. The company is developing dedicated chips to accelerate inference queries for AI models.

With demand for LLM inference surging, interest is growing in alternative architectures that do not depend on NVIDIA's GPUs. Fractile is targeting this market with an inference-focused approach. The sizeable funding round signals that investor appetite in the AI semiconductor space remains strong.

Luma Opens Up API for Image Generation Model "Uni-1.1"

Luma has started offering an API for its image generation model "**Uni-1.1**." With competitive pricing starting at $0.04 per image at 2,048-pixel resolution, it ranks third on the AI Arena leaderboard, behind Google and OpenAI.

The API includes web search, built-in reasoning capabilities, and support for up to nine reference images. With high-quality image generation now available at low cost, image generation AI is expected to become even more democratized.

Origin Lab Raises $8 Million for a Marketplace Selling Game Data for AI Training

**Origin Lab** has raised $8 million (approximately ¥1.2 billion) to build a marketplace where video game companies can sell the high-quality data they hold to AI labs.

The company sources data needed for building world models from game companies and licenses it to AI firms. High-precision 3D environment data generated by game engines is considered extremely useful for training world models in fields such as autonomous driving and robotics. As a legitimate platform for trading data, the startup aims to offer a new option for how AI training data is procured.

SenseTime Releases "SenseNova-U1," a Multimodal Model Unifying Understanding and Generation

Chinese AI company SenseTime has released the "**SenseNova-U1**" series, which unifies multimodal understanding and generation in a single architecture.

While conventional multimodal models have mostly relied on adapters to convert between modalities, SenseNova-U1 adopts a "NEO-unify" architecture that natively integrates language and vision. Its key feature is the ability to perform understanding, reasoning, and generation end to end, with the number of generation steps adjustable via LoRA. An 8B-parameter model is publicly available on Hugging Face.

Multi-Stream LLMs: Research on Improving Inference Efficiency Through Parallel Thinking Draws Attention

A paper titled "**Multi-Stream LLMs**" is gaining attention on Hugging Face's Daily Papers.

Chain-of-Thought (CoT) has significantly boosted LLM performance, but it has a drawback: representing the thinking process as a serial sequence of tokens drives up computational cost. The paper proposes processing multiple streams of thought in parallel, aiming to improve inference efficiency and break free from the natural language bottleneck. It is regarded as a highly practical approach to solving the "original sin of CoT" — the problem of forcing a single text stream to shoulder too many roles, including working memory, algorithm traces, explanations, and self-teaching targets.

---