DeepSeek V4 might finally be around the corner, and today was one of those days where AI Twitter just goes completely sideways. I covered a fresh leak showing what looks like a long list of DeepSeek V4 INT8 weight files getting uploaded—real model weight filenames, not just random “trust me bro” rumors. That matters because V4 has been the most anticipated AI release of 2026, and after weeks of missed “it’s dropping this weekend” predictions, this is the first signal that feels like someone actually touched the real thing. I also talked about the market reaction, the Nvidia shorting jokes (R1 flashbacks), and Polymarket sitting at basically a coin flip for a March release. Then it got even crazier: two stealth models appeared on OpenRouter with zero attribution—Healer Alpha and Hunter Alpha—both listed under OpenRouter’s own namespace. That’s weird because OpenRouter doesn’t build models, so someone is clearly using them as a stealth launch pad. The specs line up almost word-for-word with what’s been rumored for V4 (trillion+ MoE vibes, 1M context, agentic focus), and both pages mention prompts/completions are being logged, which screams “active development.” I ran quick tests (UI dashboard, Snake game, SaaS site) and the coding/UI output looked genuinely competitive—so if this is V4 (or V4-adjacent), it’s going to be useful fast.