Opus 4.5, Z-Image, DeepSeekMath-V2, FLUX.2, Fara-7B, Hunyuan 3D global, Supertonic TTS, ltx Retake and more

Opus 4.5, Z-Image, DeepSeekMath-V2, FLUX.2, Fara-7B, Hunyuan 3D global, Supertonic TTS, ltx Retake and more

Hey there! Welcome back to AI Brews - a concise roundup of this week's major developments in AI.

In today’s issue (Issue #115):

1. AI Pulse: Weekly News at a Glance

2. Weekly Spotlight: Noteworthy Reads and Open-source Projects

3. AI Toolbox: Product Picks of the Week

---

🗞️🗞️ Weekly News at a Glance

1. Anthropic:

   1. Claude Opus 4.5: a state-of-the art among frontier models on software coding tasks and agentic tasks that require it to run autonomously on a user’s behalf. It scored 80.9% on SWE-bench Verified with substantial improvements in reasoning, mathematics, and vision capabilities relative to earlier Claude models. Opus 4.5 demonstrates stronger prompt injection robustness in browser use than previous models [Details].

   2. Three new beta features that let Claude discover, learn, and execute tools dynamically [Details].

      

2. Black Forest Lab introduced FLUX.2 model family capable of generating highly detailed, photoreal images along with infographics with complex typography, at resolutions up to 4MP. FLUX.2 provides multi-reference support, with the ability to combine up to 10 images. It builds on a latent flow matching architecture, that unifies image generation and editing pairing a Mistral-3 24B VLM for real-world context with a rectified-flow transformer that captures spatial, material, and compositional detail beyond earlier models[Details].

   

3. Tongyi Lab Alibaba Group released Z-Image, an efficient 6-billion-parameter open-source foundation model for image generation that produces photorealistic images on par with those from models an order of magnitude larger and excels at accurately rendering complex Chinese and English text. It can run smoothly on consumer-grade graphics cards with less than 16GB of VRAM [Details].

   

4. DeepSeek released DeepSeekMath-V2 model, built on top of DeepSeek-V3.2-Exp-Base, capable of both generating and verifying mathematical proofs. It achieved gold-level performance in high-school mathematics competitions including IMO 2025 and CMO 2024. On the Putnam 2024 undergraduate competition, it scored 118/120, exceeding the highest score of 90 obtained by human participants [Details].

   

5. Microsoft released Fara-7B, an open-weight agentic SLM designed specifically for computer use. Unlike traditional chat models that generate text-based responses, Computer Use Agent (CUA) models like Fara-7B leverage computer interfaces, such as a mouse and keyboard, to complete tasks on behalf of users. With only 7 billion parameters, Fara-7B achieves state-of-the-art performance within its size class and is competitive with larger, more resource-intensive agentic systems that depend on prompting multiple large models [Details].

6. Tencent:

   1. Hunyuan 3D global: an AI platform to generate commercial-grade 3D asset production - Text-to-3D, Image-to-3D (multi-view support), or Sketch-to-3D. Achieves industry SOTA with their new 3D-DiT hierarchical carving model. The Hunyuan 3D model API is now available on Tencent Cloud International. New creators get 20 free generations daily on the platform. Enterprise API users receive 200 free credits upon registration [Details].

   2. HunyuanOCR: an open-source lightweight (1B parameters) Vision-Language Model (VLM) dedicated to OCR tasks, powered by Hunyuan’s native multimodal architecture. It achieves state-of-the-art results on OCRBench among VLMs with fewer than 3B parameters. [Details]

   3. HunyuanVideo 1.5: 8.3B parameters open-source video generation model, deployable on consumer GPUs with only 14GB VRAM. Natively generates 5–10 second 480p/720p HD videos, with super-resolution support for 1080p cinematic quality [Details].

7. ltx studio added a new feature Retake, which lets you change dialogue, pacing, framing, or emotion after a shot is rendered—without regenerating the whole video. You can instantly rephrase, reimagine, or refine specific moments while keeping the rest of the shot perfectly consistent [Details].

8. OpenAI:

   1. Shopping research: a new experience in ChatGPT that does the research for you to help you find the right products. It asks clarifying questions, researches deeply across the internet, reviews quality sources, and builds on ChatGPT’s understanding of you from past conversations and your ChatGPT memory to deliver a personalized buyer’s guide [Details].

   2. You can now use ChatGPT Voice right inside chat—no separate mode needed. You can talk, watch answers appear, review earlier messages, and see visuals like images or maps in real time.

   3. Data residency regions for ChatGPT and its API expanded letting business customers to store data in certain regions to meet local regulatory and data protection requirements [Details].

9. Perplexity:

   1. new personalization features to remember your preferences, interests, and conversations. Perplexity now synthesizes them automatically like memory, for valuable context on relevant tasks [Details].

   2. a new AI-powered shopping experience that uses conversational search and learned user preferences to provide personalized, context-aware product recommendations with integrated checkout. Available for free to all US Perplexity users [Details].

10. Character AI launched Stories, a new format where users can create and share adventures starring their favorite Characters. With Stories, users choose two or three Characters, pick a genre, and write or auto-generate a premise. From there, the experience unfolds as a guided narrative in which the user gets to make frequent choices [Details].

🔦 🔍 Weekly Spotlight

Articles/Courses/Videos:

1. Google Antigravity Exfiltrates Data - An indirect prompt injection in an implementation blog can manipulate Antigravity to invoke a malicious browser subagent in order to steal credentials and sensitive code from a user’s IDE.

2. How to Create Images &Videos with Nano Banana Pro: Complete Workflow Guide - HailuoAI

3. Effective harnesses for long-running agents - Anthropic

4. Gemini 3 Prompting: Best Practices for General Usage - Philipp Schmid

5. From shortcuts to sabotage: natural emergent misalignment from reward hacking -Anthropic

6. Claude Opus 4.5, and why evaluating new LLMs is increasingly difficult - Simon Willison

7. Claude Use Cases: Browse practical examples across research, writing, coding, analysis, and everyday tasks

Open-Source Projects:

1. Supertonic TTS: a lightning-fast, on-device text-to-speech system designed for extreme performance with minimal computational overhead. Powered by ONNX Runtime, it runs entirely on your device—no cloud, no API calls, no privacy concerns.

2. LLM Council by Andrej Karpathy: a simple, local web app that essentially looks like ChatGPT except it uses OpenRouter to send your query to multiple LLMs, it then asks them to review and rank each other’s work, and finally a Chairman LLM produces the final response.

---

🔍 🛠️ Product Picks of the Week

1. Questas: Build choose-your-own-adventure stories with AI-generated images and videos. Start with a character and a simple concept, and Questas turns it into a branching story you can edit, share, and play through like an interactive book.

2. Stanford Agentic Reviewer: Get detailed AI feedback on your research paper.

3. Google Scholar Labs: Ask a detailed research question to find relevant papers.

4. Metorial: The open source AI integration platform for developers.

Last Issue

Thanks for reading and have a nice weekend! 🎉 Mariam.