Rafomac News, Tech & Trends That MatterAI breakthroughs just keep coming. Deepseek’s new version 4 Flash Vision is now live — pushing close to the top Opus 4.8 standard. Meanwhile, Google’s Gemini 4 leaks hint at some jaw-dropping capabilities. And the mysterious OX Alpha model is quietly setting new benchmarks on Deep Sway tests, leaving the AI community buzzing.
Deepseek Version 4 Flash Vision: AI Sees Like Never Before
Deepseek’s latest release, version 4 Flash Vision Experimental, just went live on their API platform. This upgrade takes the model’s already strong text skills — including reasoning, coding, and world knowledge — and significantly boosts its visual understanding. It’s not just describing images anymore; it understands detailed visual context and integrates that with its knowledge base to complete complex tasks.
Benchmarks show that on Terminal Bench 2.1 it scored 83.9 versus Opus 4.8’s 85, while on Deep Sway it actually outperformed Opus with 59.3 to 58. The model also shines in multimodal agent evaluations, sometimes even surpassing Opus 4.8. This means it can process screenshots, interfaces, documents, and charts seamlessly, then apply tools to act intelligently — a game changer for real-world uses like visual debugging, browsing automation, and interface tasks.
A fun demo showed the AI recognizing Zank, the founder of ByteDance, from a photo. It’s a clear sign that Deepseek’s eyes are wide open now, no longer just relying on text. Developers can already access this through the newest Deepseek harness, starting practical testing right away.
Behind the Scenes: Deepseek’s Mysterious Grayscale Checkpoint
While version 4 Flash Vision gets the spotlight, Deepseek is quietly grayscale testing a mysterious new checkpoint. This experimental model is reportedly holding its own against Fable 5’s top-tier generations, creating detailed 3D builds like a helicopter simulation from a single prompt. Impressively, this autonomous task took an hour and a half but cost less than a dollar — trading time for affordability.
This grayscale testing currently routes select user sessions to the checkpoint without full public roll-out. It could be either an upgraded version 4 Pro or even version 5 in the making. The AI community is watching closely as this development hints at even more powerful models coming soon.
Gemini 4 Leaks: Google Poised for AI Comeback
Meanwhile, Google’s upcoming Gemini 4 is stirring excitement. Early leaks show it effortlessly generating detailed and coherent visual-text outputs — like a pelican riding a bike, down to perfectly rendering the tire treads and clothing folds. Few models currently reach this level of cross-modal nuance, suggesting Gemini 4 could raise the bar across multiple AI domains.
OX Alpha: The Stealth Model Breaking Records
Interest in OX Alpha keeps growing. This stealth model, free for a week with a 1 million token context window, is multimodal, data-private, and boasts an extremely high daily token capacity — reportedly up to 100 trillion tokens. New reports reveal OX Alpha scored an astonishing 80% on 10 Deep Sway tasks, outperforming Fable 5’s 65% and GBT 5.6 Soul’s 52%. One task was a close miss, meaning it might actually top 80%.
Despite initial speculation that OX Alpha could be a new GLM checkpoint, it’s now confirmed not to be GLM, Xiaomi’s Mimo, or Deepseek’s model. Rumors suggest it may come from Minimax, introducing a further twist to its origin story. Testers highlight OX Alpha’s ability to sustain long agentic runs and manage complex engineering tasks by tracking multiple files and decisions.
Creative Multimodal Magic in OX Alpha
A standout demo used Hermes Agent with OX Alpha to build a Frogger-style game. The model didn’t just replicate basic gameplay but creatively added elements like a rogue iPhone ‘enemy’ and bonus flies for the frog to collect. This ability to visually inspect and iteratively improve its outputs showcases the potential of multimodal AI agents beyond traditional coding or text generation.
Claude: Expanding Agent Abilities Beyond APIs
Not to be outdone, Anthropic’s Claude platform rolled out a major update. Its agent can now interact beyond simple API calls — using computer control to operate apps lacking APIs, navigating the web via a browser tool, leveraging reusable skills through a new Skills API, and handling documents consistently with a Files API. This suite transforms Claude into a versatile, cloud-managed AI agent system capable of complex real-world multitasking.
In related news, Codex users received a pleasant surprise with a usage limit reset following a milestone of 20 million active users, enabling continued excited experimentation through the weekend. Meanwhile, other local models like Ornith 1.5 offer impressive results, especially in lower-bit modes, showing that powerful AI isn’t limited to the biggest players alone.
These rapid developments underscore how AI is shifting fast — from better vision to creative agency to practical deployment in software and gaming. The question now is which models will define the next wave, and how fast we’ll see them in everyday tools.
