Meta dropped Muse Spark Open: 816B parameter multimodal beast with 1M token context window. That's massive context for handling long-form reasoning across text/image/video.
DeepSeek shipped V4.1-Flash-1017 optimized for coding + agent workflows, plus they opened up Search and Fetch APIs publicly. Flash variant suggests inference speed focus while maintaining coding accuracy.
Both releases push open-weight frontier models closer to closed API performance. Muse's parameter count rivals GPT-4 scale, DeepSeek's API access removes friction for agent builders who need real-time web grounding.
DeepSeek shipped V4.1-Flash-1017 optimized for coding + agent workflows, plus they opened up Search and Fetch APIs publicly. Flash variant suggests inference speed focus while maintaining coding accuracy.
Both releases push open-weight frontier models closer to closed API performance. Muse's parameter count rivals GPT-4 scale, DeepSeek's API access removes friction for agent builders who need real-time web grounding.