No GPT-6. No Gemini Ultra 3. No Anthropic surprise drop. And yet — more happened on March 22 than most people realize. Because the real story in AI right now isn't happening at OpenAI or Google. It's happening on Hugging Face, GitHub, and Cloudflare Workers — in small labs, indie teams, and research groups that ship without press releases. Here's everything that actually dropped. ## The Stealth Drop Problem The most interesting model of the past week wasn't announced. **Hunter Alpha** appeared on OpenRouter without a developer name, without a press release, without a launch event. It showed capability levels that surprised developers who tested it blind. It generated massive usage before anyone knew who built it. Then it turned out to be **MiMo-V2-Pro from Xiaomi's AI division** — built by a former DeepSeek researcher, running 1 trillion parameters, free to use. This is the new normal. Models are being tested in production, collecting real user data, before the companies behind them reveal themselves. ## Kimi K2.5 — Moonshot AI The standout release of the day. **Kimi K2.5** from Moonshot AI is now available in the Cloudflare Workers AI ecosystem — developers can run a frontier-level model directly on edge computing infrastructure. **Key specs:** - 256,000 token context window - Optimized for multi-step agentic tasks and tool calling - Strong visual understanding capabilities - Edge deployment via Cloudflare **Why it matters:** A two-person startup can now deploy the same class of model that enterprise teams were using six months ago — without managing servers. ## Qwen 3.5 Small (9B) — Alibaba Nine billion parameters. That's small by current standards — runs on a high-end laptop or recent smartphone. **The surprise:** On GPQA Diamond, Qwen 3.5 Small is matching models with 120 billion parameters. A model ten times smaller performing at the same level on complex reasoning tasks. **Practical implication:** Local AI that runs offline, without an API key, without a monthly subscription, without sending data to anyone's server. ## MiroThinker 72B — Miro Lab A 72-billion parameter model focused on reasoning and analysis tasks. Early testers report strong performance on multi-step problem solving and code generation. Details still emerging — this appears to be another stealth release following the Hunter Alpha pattern. ## Bonk — Cloudflare (Code Review Agent) Not a language model, but an **AI code review agent** built by Cloudflare. Integrates directly with development workflows to provide automated code review, security analysis, and performance suggestions. Represents a growing trend: specialized AI agents rather than general-purpose models. ## The Pattern Behind the Releases Three trends emerge from March 22's releases: 1. **Edge deployment is winning** — Kimi K2.5 on Cloudflare Workers, small models running locally 2. **Efficiency beats scale** — Qwen 3.5 Small (9B) matching 120B models 3. **Stealth launches are normal** — Models appear, get tested, then reveal their creators ## What This Means for Developers The barrier to frontier AI keeps dropping: - **Cost:** Edge deployment eliminates server management - **Privacy:** Local models don't send data anywhere - **Capability:** Small models now match yesterday's giants The companies making the most noise aren't necessarily making the most progress. ---

Find the right model for your use case at Q4KM.ai — 5,800+ models with detailed specs.