The AI and Tech Weekly: July 26, 2026

Mouad Belghiti·2026-07-268 min read

The AI and Tech Weekly: July 26, 2026

Top market news, developer tips, open source picks, and startup ideas from the community

Market News

Moonshot AI Releases Kimi K3, the Largest Open-Source Model Yet

On July 16, China's Moonshot AI launched Kimi K3, a 2.8-trillion-parameter model. The full weights go public on July 27. It is now the largest open-source model by parameter count.

The architecture uses a Mixture-of-Experts design with 896 experts. Only 16 activate per token. Two new techniques power it: Kimi Delta Attention (KDA) and Attention Residuals. Both improve how information flows across long sequences and deep layers. Compared to K2, K3 is roughly 2.5x more compute-efficient.

It ships with native vision and a 1-million-token context window. Once the weights land on July 27, expect the community to start quantizing and distilling K3 within days. Benchmarks show it matching the top closed models from OpenAI and Anthropic.

SpaceX Acquires Cursor for $60 Billion

SpaceX signed a deal to acquire Anysphere, the company behind the AI coding assistant Cursor, for $60 billion in all-stock. The announcement came on June 16. It is the largest acquisition of a venture-backed startup in history.

SpaceX had secured an option in April: pay $10 billion for a partnership or $60 billion for the full company. It chose the full acquisition. The deal is expected to close in Q3 2026, pending regulatory approval. It follows SpaceX's February merger with xAI, Elon Musk's AI venture.

Cursor had $2.6 billion in annualized revenue at the time of the deal. Around the same time, OpenAI acquired Ona, formerly Gitpod, to give its Codex agent persistent cloud sandboxes for multi-day tasks. Two major coding tools consolidated in the same month. The roadmap for Cursor now sits with SpaceX and xAI.

Together AI Raises $800 Million at an $8.3 Billion Valuation

Together AI closed an $800 million Series C on July 1. Aramco Ventures led the round. NVIDIA, General Catalyst, Salesforce Ventures, and Vista Equity Partners also joined. The new valuation of $8.3 billion is more than double the prior mark.

Together AI runs open-source models in the cloud for training, fine-tuning, and inference. Its ATLAS engine uses speculative decoding to achieve up to 400% faster inference and 60% lower costs than baseline. Annual bookings have crossed $1.15 billion. Customers include Cursor, Eleven Labs, Cognition, and Suno.

This is where capital is moving in 2026: away from model training, toward inference serving. Industry estimates put inference at roughly two-thirds of all AI compute by year-end, up from one-third in 2023. The round confirms open-source inference is now a billion-dollar market on its own.

Microsoft Launches Seven In-House MAI Models at Build 2026

At Build 2026 on June 2 and 3, Microsoft unveiled seven in-house models under the MAI brand. The lineup includes MAI-Thinking-1 for reasoning, MAI-Code-1 tuned for GitHub and VS Code, and MAI-Image-2.5-Pro for high-fidelity image generation. All seven run in production today on Bing, PowerPoint, OneDrive, Excel, GitHub Copilot, and Azure.

Microsoft says these models cut costs by up to 89% compared to OpenAI models for certain tasks. A new version of GitHub Copilot powered by MAI-Code-1 is set for general availability in August 2026. The project was developed internally under the code name Project Polaris.

Two years ago, Microsoft's AI products ran almost entirely on GPT-4. Now it has its own model stack deployed at scale across its entire product surface. This reduces dependence on OpenAI. Watch for Azure pricing changes as Microsoft passes some of those savings to enterprise customers.

Tips and Tricks from the Community

Keep your project rules in a CLAUDE.md file, not your prompts: Write project constraints, coding conventions, and repeatable workflow steps once in a CLAUDE.md file at the root of your repo. AI coding tools load it automatically at session start. You never have to repeat the same context again. (zencoder.ai/blog)
Structure coding sessions as Explore, Plan, Implement, Verify: Before asking an agent to write code, have it explore the codebase and explain the problem back to you. Then plan. Then implement. Then verify against the original spec. This four-step structure reduces errors and stops you from generating code before the problem is well-defined. (baeseokjae.github.io)
Start a fresh session for each new task: Large context windows are useful, but dumping every file into one conversation adds noise, not clarity. Use four strategies to manage project memory: write context to a persistent file, select only what is relevant per task, compress prior conversations to summaries, and isolate separate agents for separate tasks. (kunalganglani.com/blog)
Pick your coding agent based on your actual tasks, not benchmarks: No single agent wins every category. Research from mid-2026 shows tool quality depends heavily on task type. Run two or three agents on a sample of your real work before committing to a stack. Benchmark results on curated eval sets do not transfer reliably to real codebases. (developersdigest.tech/blog)

Open Source This Week

OfficeCLI: A single binary that gives AI agents full read, edit, and automation control over Word, Excel, and PowerPoint files. No Office installation required. 10,800+ stars and actively maintained as of July 6. github.com/iOfficeAI/OfficeCLI
Kimi K3 weights: Moonshot AI's 2.8T-parameter open model releases its full weights on July 27. MoE architecture, 1-million-token context, and native vision. One of the most anticipated weight drops of the year. kimi.com/blog/kimi-k3
Firecrawl: A web scraping and crawling API built to feed LLMs and AI agents with clean, structured content from any website. Used as the data layer for agents that need live web context at scale. github.com/mendableai/firecrawl
Bumblebee: A read-only supply chain scanner by Perplexity AI. It checks your project dependencies, MCP servers, and editor extensions for known-malicious packages. Reached 2,600 stars since its v0.1.1 release in May 2026. github.com/perplexity-ai/bumblebee

Startup Ideas for Inspiration

Drawn from YC's Requests for Startups and other leading accelerators

Software for Agents

Most software today was designed for humans who click buttons and scan dashboards. AI agents cannot sign up for services, fill out forms, or move through menus built for people. Integrations break. Teams spend weeks building fragile adapters before any real work gets done.

YC's Summer 2026 RFS asked founders to rebuild every major software category for a world where the primary users are AI agents, not people. The brief is specific: APIs, MCPs, CLIs, machine-readable documentation, and tools that agents can discover and start using without a human in the loop.

Pick one vertical where agents already do work today, such as accounting, legal research, or project tracking. Build an API-first product with full MCP support. Write documentation that an agent can parse and act on in one pass. Price per API call or per task completed, not per seat.

AI-Native Service Companies

Professional service firms like agencies, consultancies, and legal shops charge high rates for work that is increasingly automatable. A 10-person agency often cannot compete on price with a two-person team using agents doing the same output.

YC's RFS frames this as one of the biggest gaps in the market today. Service businesses are worth hundreds of billions globally. The first firms to go fully AI-native can price at a fraction of incumbent rates while keeping better margins. The window before incumbents react is open right now.

Pick a narrow service with well-defined outputs: SEO audits, contract review, or financial modeling. Productize it so an agent handles 90% of the work. Charge per output, not per hour. Start with 5 to 10 clients and build only the software those clients actually need.

Dynamic Software Interfaces

Current software has fixed interfaces built for an average user. An agent that has used your product for three months sees the same forms and menus as a first-time visitor. Nothing adapts. Every session starts from scratch with the same 10-step wizard.

YC sees dynamic interfaces as a multi-billion-dollar gap. Software that builds its own interface on the fly, based on who or what is using it, changes the economics of software products. An agent should call a tool, not click through a UI designed for someone who has never seen the product before.

Pick a category with repetitive, predictable tasks: CRM entry, HR onboarding, or invoice processing. Build a thin UI layer that tracks how specific users complete tasks and collapses the steps they never use. Then add a machine-readable API layer that agents use directly, bypassing the UI entirely.

The AI OS for Companies

Most companies store knowledge across Slack, Notion, email, and GitHub with no single source of truth. New employees spend weeks getting up to speed. AI agents running multi-step workflows have no reliable place to look up basic company context. They fill in gaps with guesses.

YC calls this the "Company Brain." Enterprises are now deploying agentic workflows that run for hours or days to handle end-to-end business functions. These agents need persistent, structured access to company knowledge. No major platform has fully solved this, and the market is large.

Start with one engineering team of 10 to 20 people. Build a structured memory layer that indexes their Slack, GitHub, and docs. Expose an API that coding agents can query mid-task. Charge per team per month. Once the pattern works for one team, it expands company-wide with minimal extra engineering.

※