Transcription
[Music] Something really strange is happening right now in the AI world. On one side, a mysterious model called Horizon Alpha suddenly showed up on Open Router without a single announcement, no author, no documentation, just anonymous slapped on it. And on the other side, some suspicious GitHub repositories briefly went live with names like Yofo Wildflower and Yofo Deepcurren containing what looked like configs for massive open-source GPT style models. People spotted those before they vanished, and the connection between both events is way too tight to be a coincidence.
Let's start with Horizon Alpha. It was dropped quietly on July 31st and within hours it climbed to the top of EQBench, the benchmark known for testing creative reasoning, emotional intelligence, and long-form writing. These aren't raw math tests or factual recall. This is where most models struggle to sound human, stay coherent over multiple paragraphs, or maintain subtle narrative flow. Horizon Alpha didn't just compete, it destroyed everything in its path.
Now, here's where it gets weird. Just as Horizon Alpha was making noise, the AI community started uncovering leaked model repositories tied to open AI staff accounts. One of them was labeled Yofo Wildflower/GPTOSS20B and the other YOFO deepcurren/poss 120B. That GPTOSS tag pretty clear it stands for GPT open-source software. And those two model names, Wildflower and Deep Current, seem to represent a small and large version of the same architecture. The timing of these leaks couldn't have been worse for secrecy and better for speculation.
Because Horizon Alpha, it fits, it behaves like a highly capable base model. It's extremely fast, spitting out roughly 150 tokens per second. It has a 256,000 token context window. That's huge. It reads images, interprets complex puzzles, writes clean HTML visualizations for spatial logic problems. One person gave it a task from a kid's picture book just saying, "Read the text and do what it says." And it did flawlessly. OCR, reasoning, vision, all working together.
Now, combine that with what was found in the leaked config from the YFO deepurren repo. That model, the suspected 120 billion parameter GPTO OSS is structured as a mixture of experts. This means that while the full model holds 120 billion parameters, only a small group about 5 billion are activated per query. This makes the model fast, memory efficient, and shockingly cheap to run considering its size, which would explain why Horizon Alpha runs as fast as it does. It acts like a refined architecture and just like the models in those repos, Horizon Alpha lacks obvious safety alignment. It agrees with almost anything, doesn't challenge bad ideas and fails at simple math logic traps. Those are usually the first things fixed in a commercial release. But if it's a base model or an open-source drop being tested publicly, it starts to make sense. You don't safety align until after the base is finalized.
Also, get this. When Horizon Alpha was asked directly who made it, it responded, "I'm an Open AI language model GPT4 class. I was created by Open AI." It didn't hesitate. That triggered a wave of Reddit threads asking if OpenAI is quietly testing GPT5 capabilities under the radar. The theory is that Horizon Alpha might be a stripped down test version of GPT5 or an openweight sibling with different tuning and the GitHub leaks. They might have revealed the open source plan behind it.
Let's go back to those for a second. The repositories weren't empty. One of them included a full config for a model with mixture of experts, a giant vocabulary set, and support for sliding window attention. That attention mechanism allows the model to handle very long sequences of text without performance loss, which matches Horizon Alpha's huge context length. And inside those configs, people noticed something unusual. FP4 precision 4bit floatingpoint weights. That's half the size of FP8 and a quarter of FP16. If true, that would make GPT OSS120B one of the most memory efficient large models ever built. Instead of needing 240 GB of VRAM to load, it could run with just 60. In theory, a high-end gaming PC or workstation with some RAM headroom could actually load and run this monster locally if the inference is optimized.
People started asking, "Wait a second, is OpenAI testing FP4 training here?" That's not just compression. Training directly in FP4 is hard. You lose so much numerical precision. The gradients go wild and models collapse unless your training process is rock solid. But if they pulled it off, it's a breakthrough. It opens the door to training huge models with less compute and running them on far smaller machines.
Now, not everyone's convinced. Some argue it's more likely that the model was trained in FP16 then quantized down to FP4 afterward. That's still impressive, but it's not the same thing. However, there's no quantization config in the leaked files and no mention of post-processing steps. That absence is exactly what's fueling the theory that FP4 might have been used from the beginning. If true, we're looking at a turning point in model training efficiency.
Meanwhile, the model's name, Horizon Alpha, might just be a placeholder. Or maybe it's symbolic, a new horizon, a starting point, alpha version. What's even more telling is how quickly people tied it to Open AI. There's a reason for that. It's not just the benchmark dominance or the language style. It's the fact that Open AI has been under pressure lately. Pressure from competitors and internally. The collapse of their $3 billion deal to acquire Windsurf made headlines. First, Anthropic pulled their models from the partnership. Then, Microsoft, yes, their closest ally, reportedly blocked the acquisition to protect GitHub co-pilot. And just when things couldn't get messier, Google stepped in and hired Windsurf's top engineers. That left OpenAI with nothing but a PR headache and no strategic win.
Add to that a rumored restructuring plan to turn OpenAI into a fully for-profit company to raise $40 billion tied to a clause with a 20 billion penalty if certain milestones aren't met. That kind of money pressure means they have to ship something massive, something like GPT5 or a family of open-source models that can recapture developer goodwill. Because while Open AI has been dealing with internal chaos, their competitors haven't slowed down. China is moving fast. Alibaba launched Quen 3 thinking, a model that outperforms OpenAI and Google on reasoning and code generation benchmarks. Moonshot AI released a 1 trillion parameter agentic model. Z.AI dropped GLM4.5 which now ranks third overall across all proprietary and open models. And over in Europe, Mistl has been building compact highquality openweight models designed for local inference. Their defro model is optimized for coding tasks and runs on consumer hardware. These companies aren't just experimenting, they're releasing polished, aligned, openweight models with clear use cases. That's a lot of pressure on Open AI, especially if they want to stay in front.
So, what happens now? Horizon Alpha continues to sit at the top of the EQ bench leaderboard. Developers keep pushing it, testing its limits, and trying to connect the dots. At the same time, the community waits to see if OpenAI will officially release the models linked to YOFO Wildflower and Yofo Deep Current. Will we see a 20 billion parameter and a 120 billion parameter GPTO OS family released on HugenFace or Open Router? Or was this just a tease to test the water? And let's not forget, some people believe Horizon Alpha and GPTO OS are part of the same plan. Horizon Alpha could be the aligned test bed for creative tasks, while GPTOS might be the stripped down efficient version for developers and open-source enthusiasts. Different rappers, same core, or maybe not. Maybe Horizon Alpha really is a cloaked GPT5 running under a generic name to gather real world feedback before a full reveal.
Whatever it is, the fact that it can generate long coherent stories, solve spatial puzzles, understand image text combinations, and respond with zero delay, all without an official identity is kind of insane. One way or another, the world's most famous AI lab just dropped a model that doesn't admit it exists. And right next to it, we've got deleted GitHub repos leaking what looks like the blueprint for a massive open-source strategy. Whatever OpenAI is doing, they've made one thing clear. Something big is coming.
So, what do you think? Is this OpenAI's way of going open source, or are they hiding something bigger? Drop your thoughts in the comments. I read them all. Don't forget to subscribe and hit the like button if you found this interesting. Thanks for watching and I'll catch you in the next one.