AI Dev Weekly is a Thursday series where I cover the week’s most important AI developer news, with my take as someone who actually uses these tools daily.

Microsoft Build 2026 was the story of the week. Seven in-house AI models — none trained on OpenAI data. A Surface mini PC with NVIDIA RTX Spark inside. Claude Code licenses cancelled, developers pushed to Copilot. This was Microsoft saying, loud and clear: we don’t need OpenAI anymore. Meanwhile MiniMax dropped the first open-weight frontier multimodal model, and NVIDIA unveiled hardware that makes local AI actually practical.

1. Microsoft Build 2026: 7 models, zero OpenAI dependency

Microsoft unveiled its MAI (Microsoft AI) model family at Build on June 2. The headline: MAI-Thinking-1 — a 35B reasoning model trained entirely on commercially licensed enterprise data with no distillation from GPT or any OpenAI model. It matches Claude Sonnet 4.6 on key benchmarks at up to 10× better cost efficiency.

Other MAI models announced:

• MAI-Code-1-Flash (5B) — Purpose-built for GitHub Copilot and VS Code
• MAI-Thinking-1 (35B) — Reasoning, multi-step instructions, long context
• Aion 1.0 Instruct — Local Windows model for on-device reasoning
• Aion 1.0 Plan — Local Windows model for planning and tool use
• Plus 3 more across transcription, speech, and images

Also announced: Windows as “agent-native runtime” with Microsoft Execution Containers (MXC) — sandboxed environments for running AI agents with enterprise-grade isolation.

My take: This is the biggest signal yet that the OpenAI-Microsoft marriage is evolving into a polite separation. MAI-Thinking-1 being trained without any OpenAI data is deliberate positioning — Microsoft can now say “our models have no license entanglement with OpenAI.” The 5B coding model (MAI-Code-1-Flash) specifically targets Copilot — Microsoft’s most important developer product. They’re replacing GPT inside their own tools with models they fully control.

2. Microsoft ends Claude Code licenses

Forbes reported that Microsoft is ending Claude Code licenses and pushing developers to its own Copilot CLI instead. The subtext: Microsoft no longer wants to rent Anthropic’s intelligence inside its own products.

This affects developers at Microsoft who were using Claude Code as their primary coding tool. They’re being migrated to Copilot powered by MAI-Code-1-Flash.

My take: If you work at a Microsoft shop: prepare for Copilot to get much better (MAI-Code-1 is purpose-built for it). If you use Claude Code independently: nothing changes for you. But this signals that the era of “one AI provider to rule them all” is over. Every major tech company is building their own models now.

3. NVIDIA RTX Spark: the hardware that makes local AI real

At Computex (June 1), NVIDIA unveiled RTX Spark — a new Windows PC superchip with 128GB unified memory, ARM CPU, Blackwell GPU, 1 petaflop of AI compute. It runs 120B parameter models locally.

Then at Microsoft Build (June 2), Microsoft announced the Surface RTX Spark Dev Box — a mini PC with RTX Spark inside, preloaded with VS Code, GitHub Copilot, WSL2 with GPU passthrough, CUDA, Python, Git, and Node.js. It’s purpose-built for AI developers.

Key numbers:

• 128GB unified memory (run Qwen 3.6 27B at 2× speed, Llama 4 Scout, etc.)
• 100W sustained thermal design in aluminium chassis
• Ships with Windows 11 Pro + full dev stack preinstalled
• Available this fall alongside consumer RTX Spark laptops

My take: This is the machine I’ve been wanting for the AI Startup Race. Currently our agents run on a $40/mo VPS. The Surface RTX Spark Dev Box would let you run 120B models locally with zero API costs. For developers spending $100+/month on AI APIs, this hardware pays for itself fast. See our full RTX Spark guide and vs Mac Studio comparison.

4. MiniMax M3: first open-weight frontier multimodal

MiniMax M3 launched June 1. First open-weight model combining:

• 59% SWE-bench Pro (beats GPT-5.5’s 58.6%)
• 1M token context via MSA (15.6× faster than standard attention)
• Native text + images + video input
• Computer use (desktop operation)
• $0.60/$2.40 per million tokens
• Weights dropping ~June 10

It reproduced an ICLR 2025 paper autonomously — 12 hours, 18 commits, 23 figures, zero human intervention.

My take: M3 is the model that should worry Anthropic most. It beats GPT-5.5 on coding while being open-weight, multimodal, and 10× cheaper than Opus. When weights drop (~June 10), enterprises with data privacy requirements get a frontier model they can run on-premise. That was previously impossible without Claude’s closed API. See our M3 complete guide, vs Claude Opus 4.8, and vs DeepSeek V4-Pro.

5. Grok’s next model teased (mid-June)

xAI teased their next Grok model — reportedly tripling the parameter count with a focus on coding leadership. Expected release: mid-June 2026. Supervised fine-tuning was complete as of the announcement, with reinforcement learning underway.

If it delivers on the coding promise, it could shake up the Grok Build ecosystem significantly. Currently Grok Build uses Grok 4.3 — an upgrade to whatever this new model is could make the $30/mo subscription much more competitive.

6. Quick hits

• OpenAI Codex on Amazon Bedrock — GPT models + Codex now generally available on AWS with enterprise controls. OpenAI meeting customers where they already deploy.
• White House signed frontier-model cyber order — Regulation incoming for the most capable AI models used in cybersecurity applications.
• Anthropic expanded Project Glasswing — More organizations getting access to Mythos Preview for cybersecurity work. Broader release “in weeks.”
• NVIDIA Nemotron 3 Ultra free on OpenRouter — NVIDIA’s own model now available at zero cost via OpenRouter.

What I’m watching next week

• MiniMax M3 weights (~June 10) — Will they live up to the benchmarks when the community tests them?
• Grok new model (mid-June) — Does tripling parameters actually improve coding?
• Gemini CLI sunset (June 18) — Two weeks left. Migrate now if you haven’t.
• Anthropic Mythos — Still “coming in weeks.” Will it ship before the Grok model does?
• Our race agents — Xiaomi hit session 456. 605 users/week on its site. Zero revenue still. 5 weeks left. Follow along →

---

AI Dev Weekly publishes every Thursday. Subscribe for weekly race updates and AI developer news.