⚡AI Agents Weekly: Mistral Agents API, FLUX.1 Kontext, DeepSeek-R1 Update, Codestral Embed, AgentSeek
Mistral Agents API, FLUX.1 Kontext, DeepSeek-R1 Update, Codestral Embed, AgentSeek
In today’s issue:
FLUX.1 Kontext
DeepSeek-R1-0528
Mistral AI Agents API
Google releases MedGemma models
Mistral AI has released Codestral Embed
Anthropic open-sourced a new interpretability toolkit
AgentSeek is a local alternative to Manus AI
Building production-grade conversational agents with DAG
An operating system for memory-augmented generation in LLMs
Top AI dev news, AI agents research, and much more.
Top Stories
FLUX.1 Kontext
Black Forest Labs introduces FLUX.1 Kontext, a suite of multimodal flow-based models for image generation and editing that blend text and visual inputs to produce high-quality, contextually coherent outputs. Unlike traditional text-to-image models, FLUX.1 Kontext enables in-context generation, preserving characters, styles, and local details across edits, while also supporting rapid, iterative transformations through natural language instructions.
Key highlights:
Unified editing and generation: FLUX.1 Kontext merges image editing and generation into a single model pipeline. Users can start with an image and use text prompts to perform local or global changes, or initiate generation purely from text, with strong prompt adherence and photorealism.
Fast and interactive: The FLUX.1 Kontext [pro] model is optimized for speed and iteration, enabling multi-step edits with minimal latency while maintaining consistency in characters and style. The experimental [max] variant improves prompt following and typography without sacrificing speed.
Context-aware capabilities: It supports local edits, character consistency across scenes, and the ability to transfer styles from reference images. Users can iteratively build complex scenes, e.g., “make her laugh,” then “change setting to a nightclub”, while preserving visual coherence.
Strong benchmark performance: On the new KontextBench benchmark, FLUX.1 Kontext [pro] leads in text editing and character preservation, while outperforming state-of-the-art models in both quality and inference latency.
Open and extensible: The [dev] model variant, released in private beta with open weights, allows researchers to explore and customize a 12B diffusion transformer compatible with earlier FLUX infrastructure.
DeepSeek-R1-0528
DeepSeek has announced a new update to their DeepSeek R1 model, with improvements in benchmark results, enhanced front-end capabilities, reduced hallucinations, and support for JSON output and function calling.
DeepSeek also announced DeepSeek-V3-0324 with a boost in reasoning performance, stronger front-end capabilities, and smarter tool use.
DeepSeek’s R1-0528 update jumps to #2 globally in the Artificial Analysis Intelligence Index, matching Gemini 2.5 Pro and outperforming Meta, xAI, and Anthropic. Achieving major gains in math, coding, and reasoning through reinforcement learning alone, R1 proves open-weight models can now rival top proprietary labs.
Keep reading with a 7-day free trial
Subscribe to AI Newsletter to keep reading this post and get 7 days of free access to the full post archives.