Higgsfield Launches Adobe Firefly Rival

In this newsletter, we spotlight three groundbreaking innovations redefining how AI intersects with creativity and software engineering:

🎨 Higgsfield Soul Inpaint brings pixel-perfect image editing with a distinct artistic flair, allowing creators to modify elements like hair, clothes, or objects while maintaining the signature Soul aesthetic. Just retweet the official post to unlock a full usage guide.

👨‍💻 DeepSWE-Preview sets a new benchmark for open-source coding agents. Trained solely through reinforcement learning, this powerful model demonstrates real-world coding prowess—handling bug fixes, feature implementations, and debugging—without any supervised fine-tuning.

⚡ Osmosis-Apply-1.7B proves that smaller can be smarter. This lightweight, RL-fine-tuned model outperforms larger foundation models in code merging tasks, offering low-latency, high-accuracy performance—perfect for seamless IDE integration.

Let’s dive deeper into the future of creative precision and efficient software development.

Higgsfield Soul Inpaint: Precision Edits with Signature Soul Style

Higgsfield Soul Inpaint is an advanced AI-powered tool that enables users to perform pixel-perfect inpainting while preserving the distinctive Soul-style high aesthetic. With this tool, you can seamlessly edit or modify specific elements in your images—such as clothes, hair, or objects—without losing the unique artistic quality that defines the Soul style. Designed for creative flexibility and high-quality results, Soul Inpaint is ideal for anyone seeking detailed and precise modifications in their visual projects. To access a comprehensive guide on how to use the tool, simply retweet the official announcement post, and you will receive the full instructions via direct message.

DeepSWE-Preview: Open-Source RL-Powered Coding Agent

DeepSWE-Preview is a fully open-sourced, state-of-the-art coding agent developed through a collaboration between the Agentica team and Together AI, trained entirely from scratch on top of Qwen3-32B using only reinforcement learning (RL). Leveraging Agentica’s rLLM post-training framework, DeepSWE-Preview was trained over six days on 64 H100 GPUs with 4,500 real-world software engineering tasks from the R2E-Gym benchmark, focusing on complex, multi-step tasks like bug fixing, code feature implementation, and debugging. The agent achieved a remarkable 59% accuracy on the SWE-Bench-Verified benchmark with test-time scaling (42.2% Pass@1, 71.0% Pass@16), outperforming all previous open-weight coding agents. Notably, DeepSWE-Preview demonstrates emergent behaviors such as reasoning about edge cases, allocating thinking tokens based on step complexity, and self-verifying against regression test suites—all arising purely from RL training without supervised fine-tuning or distillation. The entire training recipe, codebase, dataset, and evaluation logs are open-sourced, enabling the community to reproduce results and further advance RL-based agentic AI for software engineering.

Osmosis-Apply-1.7B: Fast RL Model for Code Merging

Osmosis-Apply-1.7B is a compact, open-source language model fine-tuned with reinforcement learning on Qwen3-1.7B to excel at code merging tasks, offering a faster and more efficient alternative to large foundation models like OpenAI o3, Claude 4 Sonnet, and Gemini 2.5 Flash. Trained on 100,000 code commit examples from the CommitPackFT dataset using a simple reward system—full reward for perfect merges, partial reward for minor formatting issues, and none for failures—the model achieved a near-perfect reward score of 0.9893, outperforming all tested foundation models while running at much lower latency and cost. Osmosis-Apply-1.7B is designed for direct integration into AI IDEs via the Model Context Protocol (MCP) server, enabling local, in-place code editing and merging with high accuracy and efficiency. The model’s training approach focused on practical efficiency, using GRPO, minimal regularization, and a single epoch of training across 8 GPUs, demonstrating that a well-designed RL reward function can yield strong results even with limited data and compute.

Automate your LinkedIn and social media growth with Olly.social — your AI assistant that brings real engagement, real reach, and real results.

Top AI Products from this week

• PromptForge - AI prompt engineering workbench for crafting, testing, and systematically evaluating prompts with powerful analysis tools. - insaaniManav/prompt-forge.

• Buildrs - The ultimate platform for vibe coders and creative builders. Connect with developers who code with passion, build amazing projects, and create the future. Find your coding tribe among the most talented vibe coders worldwide.

• AI Mindmap Extension - Turn web pages, PDF, YouTube videos, and AI chats into clear, visual mind maps with one click. Enhance learning, extract key insights, and organize information seamlessly. Integrate with FunBlocks AIFlow for deeper exploration and AI-powered brainstorming.

• Kyutai TTS - Kyutai TTS is a new open-source text-to-speech model optimized for real-time use. It's the first TTS that streams text in as it streams audio out, enabling ultra-low latency for LLM applications.

• Exla FLOPs - Exla FLOPs is the only service where you can instantly spin up 64, 128, or more GPUs - no waitlists, no commitments. Just clusters at your command.

• LumozAI - LumozAI is your all-in-one AI hub to access, deploy, and build with powerful models. Use ready-made APIs, deploy LoRAs, and soon train your own models. Easily create custom apps with our platform—AI integration has never been this simple.

This week in AI

• AI + Underwater Photography - MIT’s LOBSTgER project fuses generative AI with underwater photography, creating scientifically accurate, immersive images of marine life in the rapidly changing Gulf of Maine—expanding how we visualize and communicate ocean biodiversity.

• H&M CONTINUES ITS EXPLORATION OF CREATIVITY WITH AI - H&M continues its exploration within generative AI by dropping its first set of images featuring digital twins, set to the background of fashion capitals around the world, showcasing seasonal yet timeless denim.

• Google Gems in Workspace Side Panel - Gems, custom AI experts, are now in the side panel of Google Docs, Sheets, Slides, Drive, and Gmail. Use or create Gems to boost productivity without switching apps.

• EU Sticks to AI Act Timeline - The EU will enforce AI Act deadlines as planned: obligations for general-purpose AI models start August 2025, high-risk models in August 2026, despite calls for delay.

• AI-Written Papers Surge - Study finds 13.5–40% of biomedical abstracts may be AI-generated, suggesting at least 200,000 papers yearly use AI, raising concerns over scientific integrity.