DeepSeek strikes again, new top image models, Claude Opus 4.5, open source robots: AI NEWS

Key Concepts

• OCR (Optical Character Recognition): Technology that converts images of text into machine-readable text.
• AI Agent: An AI system capable of autonomously performing tasks and making decisions.
• Open-Source: Software or models whose source code is made publicly available for use, modification, and distribution.
• Parameters: In AI models, parameters are the variables that the model learns during training. More parameters generally mean a more complex and potentially more capable model.
• VRAM (Video Random Access Memory): Memory on a graphics card (GPU) used for storing data and computations related to graphics and AI.
• JSON (JavaScript Object Notation): A lightweight data-interchange format that is easy for humans to read and write and easy for machines to parse and generate.
• Vision-Language-Action (VLA) Model: An AI model that integrates visual understanding, language comprehension, and the ability to perform actions.
• Reinforcement Learning: A type of machine learning where an agent learns to make decisions by taking actions in an environment to maximize a cumulative reward.
• Imitation Learning: A machine learning technique where an agent learns to perform a task by observing and mimicking human demonstrations.
• Teleoperation: Remote control of a robot or device by a human operator.
• Context Window: The amount of text or data an AI model can consider at one time when processing a prompt or generating a response.
• Prompt Injection: A security vulnerability where malicious input is inserted into a prompt to manipulate an AI model's behavior.
• Quantization: A technique used to reduce the size of AI models by representing their parameters with lower precision numbers.
• NPU (Neural Processing Unit): A specialized processor designed to accelerate AI and machine learning tasks.

New AI Image Generators

This week saw the release of several new AI image generators, with a focus on realism, detail, and editing capabilities.

Flux 2

• Capabilities: Generates highly realistic and detailed images up to 4 megapixels. It can also edit existing images, transfer styles, and add elements while preserving character consistency. It supports up to 10 reference images for generation.
• Versions:
  • Flux 2 Pro: Offers the best quality but is closed-source and paid.
  • Flux 2 Dev: An open-source version with significantly lower quality, described as having a "fake plastic vibe." It's a large 32 billion parameter model requiring substantial VRAM (at least 64 GB with offloading) and is slow to generate images.
  • Flux 2 Klein: An upcoming smaller, distilled open-source version with an Apache 2 license, suitable for commercial use.
• Critique: The speaker found Flux 2 Pro to be inferior to Nano Banana Pro (released the previous week) and Flux 2 Dev to be worse than Alibaba's Quen Image. The high resource requirements and lower quality of the open-source version make it unappealing.

Zimage

• Capabilities: An open-source image generator and editor that is significantly better than Flux 2 and much smaller. It excels at realistic image generation, rendering text accurately, and handling human anatomy. It also supports uncensored content.
• Models:
  • Zimage Turbo: A tiny model (6 billion parameters) that fits comfortably within 16 GB of VRAM and runs very fast, generating images in seconds. Quantized versions are available for even lower VRAM.
  • Zimage Base: The base model, with checkpoints planned for release to the open-source community for fine-tuning.
  • Zimage Edit: An upcoming model for editing existing images, with capabilities like adding images to galleries, changing elements (text, objects), modifying character appearance, and changing artistic styles or perspectives.
• Significance: Considered one of the best open-source models currently available. A full installation tutorial and review are planned.

I Montage

• Capabilities: A versatile image generation and editing tool that can handle one or multiple input images and output one or multiple images.
  • Editing: Edits existing photos with text prompts, changing backgrounds, colors, adding elements, removing objects, altering materials, and adjusting facial expressions. It can preserve character consistency across edits.
  • Multiple Inputs/Outputs: Can combine multiple input images or generate multiple consistent output images from a single input.
  • ControlNet Integration: Features built-in ControlNet functionality, allowing control over composition using depth maps, pose maps, or edge maps.
  • Style Transfer: Can transform one photo into another style using a reference image.
  • Perspective/Angle Changes: Can alter the angle or perspective of a photo, simulating movement or changes in viewpoint.
  • Consistent Storyboards: Capable of generating multiple consistent output photos to create storyboards with consistent characters, objects, and backgrounds.
• Technical Details: The model is approximately 26 GB in size, suggesting it can be run with 16-24 GB of VRAM with offloading.
• Availability: Released with instructions on GitHub for local setup.

AI Agents and Automation

Several new AI agents and models were introduced for autonomous computer operation and task execution.

Hunyen OCR (Tencent)

• Capabilities: An AI model for understanding and parsing text within images with state-of-the-art performance.
• Key Features:
  • Accurately parses complex tables from academic papers.
  • Extracts information from invoices into structured JSON objects.
  • Parses and reformats complex charts.
  • Recognizes chemical formulas and challenging handwriting styles.
• Technical Details: A remarkably tiny model with only 1 billion parameters, yet it outperforms larger proprietary models like Gemini 2.5 Pro and GPT40, and even the recent DeepSeek OCR.
• Availability: Open-source, with instructions for local download and execution on GitHub. Requires a CUDA GPU with at least 20 GB of VRAM.

Geo Vista

• Capabilities: An AI agent that accurately determines the location where a photo was taken.
• Methodology: Autonomously analyzes images, parses text, zooms in for clues, and performs web searches. It can understand and process text in different languages.
• Performance: Outperforms open-source alternatives and is competitive with top closed-source models on various benchmarks.
• Technical Details: A 7 billion parameter model, approximately 33 GB in size, runnable on high-end consumer GPUs.
• Availability: Open-source, with setup instructions on GitHub and a Hugging Face repository.

Farra 7B (Microsoft)

• Capabilities: A tiny, open-source agentic model designed for autonomous computer operation.
• Functionality: Can perform tasks like shopping, booking travel, searching for information, and filling out forms. It can see the screen (using a vision model like Quen 2.5VL) and control the mouse and keyboard.
• Key Features:
  • Operates autonomously, mimicking human computer interaction.
  • Stops at critical points for human input or approval.
  • Can perform multi-step tasks like finding and summarizing online information.
• Performance: More performant and cost-efficient than other computer use agents, including UI tars and OpenAI's computer use solutions.
• Technical Details: A 7 billion parameter model, approximately 16 GB in size. It can run on consumer devices and is optimized for NPU acceleration on Copilot+ PCs, not just Nvidia GPUs.
• Availability: Open-source under the MIT license.

Ry VLA2

• Capabilities: A unified vision, language, action, and world model that combines these elements to control robots.
• Functionality: Can be prompted to perform tasks like picking up objects and placing them in designated locations. It can distinguish between different items, adapt to changes in the scene (additional items, moving objects, camera obstructions, height variations), and carry out tasks correctly.
• Availability: Open-source under the Apache 2 license, with instructions for installation and execution on GitHub.

ChatGPT Shopping Research

• Capabilities: An autonomous agent integrated into ChatGPT that assists users with product research.
• Functionality: Users describe their needs, and ChatGPT asks clarifying questions, researches online from trusted sources, and provides a personalized buyer's guide.
• Key Features:
  • Synthesizes information from multiple retailers.
  • Provides top options, clear differences, and trade-offs.
  • Interactive refinement process (like Tinder) to understand user preferences.
  • Ensures privacy as chats are not shared with retailers.
  • Results are organic and based on publicly available retail sites.
• Technical Details: Powered by a specialized version of GPT-5 Mini, fine-tuned for shopping tasks.
• Availability: Rolled out to all ChatGPT users, including those on the free plan, with near-unlimited usage during the holidays.

Advanced AI Models

Significant advancements were made in specialized AI models for mathematics and coding.

Deepseek Math V2

• Capabilities: A specialized AI model for advanced mathematical reasoning.
• Achievements: Achieved gold medal status on the International Math Olympiad (IMO) 2025 and Canadian Math Olympiad (CMO) 2024, and a near-perfect score on the Putnam 2024 competition.
• Methodology: Built on Deepseek V3.2, it uses a unique approach of training a verifier to check the correctness of each reasoning step, rather than just rewarding the final answer. This "self-verification" method is crucial for complex mathematical tasks requiring rigorous step-by-step derivation.
• Performance: Achieved 99% on a benchmark and scored very close to Gemini Deep Think on Proofbench Advanced, outperforming proprietary models like Gemini 2.5 Pro and GPT40.
• Availability: Open-source under the Apache 2 license.

Claude Opus 4.5 (Anthropic)

• Claimed Strengths: Touted as the best model in the world for coding and agentic use.
• Performance Metrics:
  • Achieved 80.9% on the SUIB bench verified benchmark for software engineering tasks, outperforming previous Claude models, Gemini 3 Pro, and GPT 5.1 Codeex Max.
  • Claims to write better code across seven out of eight programming languages.
  • Solves problems in fewer steps, using fewer tokens.
  • More robust against prompt injection attacks.
• Independent Benchmarks & Critiques:
  • Overall intelligence rankings place it tied for second with GPT 5.1, behind Gemini 3 Pro on the Artificial Analysis leaderboard.
  • Context Window: Only 200,000 tokens, significantly smaller than Gemini 3 Pro's 1 million tokens, limiting the amount of information that can be processed at once.
  • Price: Extremely expensive at $10 per million tokens, twice as expensive as Gemini 3 Pro.
  • ARC AGI 2 Benchmark (Visual Puzzles/Learning): Gemini 3 scores higher, and Gemini 3 Deep Think scores even higher.
  • LM Arena (Text Chatting): Ranks third, below Grok 4.1.
  • SUIB Benchmark: Only 2% better than Gemini 3 Pro on the official leaderboard, despite being much more expensive and having a smaller context window.
• Conclusion: While strong in agentic coding, it's not as well-rounded as Gemini 3 Pro and is significantly more expensive. The speaker was not overly impressed and suggests waiting for future releases from OpenAI, XAI, and Google.

Robotics Advancements

Progress was highlighted in humanoid robotics and affordable home robots.

Unree G1

• Capabilities: A highly flexible and acrobatic humanoid robot.
• New Demo: Showcased autonomously playing basketball, demonstrating smooth, natural movements, pivoting, shooting over opponents, and dribbling with fast hand-eye coordination and balance.
• Significance: Demonstrates impressive advancements in training robots for complex physical tasks.

Aloha Mini

• Capabilities: An open-source home robot designed for household chores.
• Key Features:
  • Affordable: Costs around $600 to build using a 3D printer and readily available components.
  • Dual-armed with wheels.
  • Learns tasks through teleoperation and imitation learning.
  • Can perform tasks like picking up items, wiping tables, opening fridges, and putting clothes in laundry baskets.
• Technical Details: Components include a Raspberry Pi, five cameras, motors, and a battery. Assembly takes about 60 minutes.
• Availability: Completely open-source, with hardware requirements, assembly guides, and software setup instructions available.

Conclusion

The AI landscape continues to evolve at an unprecedented pace, with significant breakthroughs in image generation, autonomous agents, specialized models for complex reasoning, and robotics. While proprietary models push boundaries, the open-source community is rapidly catching up, offering powerful and accessible tools. The focus is shifting towards more efficient, specialized, and user-friendly AI applications, with a growing emphasis on practical real-world deployment and affordability. The rapid iteration cycle suggests that even more advanced capabilities are on the horizon.