DeepSeek V4, SenseTime U1, and Poolside Laguna Launch in Major AI

Multiple AI companies released significant new models this week, with DeepSeek’s V4 flagship, SenseTime’s U1 image processor, and Poolside’s Laguna coding models marking a surge in open-source AI development. The releases span different capabilities — from long-context processing to image understanding and agentic coding — while highlighting the growing competition between Chinese and US AI firms.

DeepSeek Returns with V4 Flagship Model

Chinese AI firm DeepSeek on Friday released a preview of V4, its most significant model since the breakthrough R1 reasoning model launched in January 2025. According to MIT Technology Review, V4 can process much longer prompts than previous generations through a new design that handles large amounts of text more efficiently.

The V4 release comes after months of scrutiny for DeepSeek, including major personnel departures and delays to previous model launches. The company had maintained a relatively low profile since R1 stunned the global AI industry with its strong performance despite being trained on limited computing resources.

Like all DeepSeek models, V4 is open source and available for anyone to download, use, and modify. The company effectively teased the release earlier this month by adding “expert” and “flash” modes to its online model interface, prompting speculation about a bigger upcoming announcement.

SenseTime Launches Speed-Focused U1 Image Model

Sanctioned Chinese AI company SenseTime on Tuesday released SenseNova U1, an open-source model that can generate and interpret images without first translating them to text. According to Wired, this approach significantly speeds up processing and reduces computing power requirements compared to competing US models.

“The model’s entire reasoning process is no longer limited to text. It can reason with images as well,” Dahua Lin, SenseTime’s cofounder and chief scientist, told Wired. Lin, also a professor at the Chinese University of Hong Kong, said models capable of direct image processing will enable robots to better understand the physical world.

SenseTime designed U1 to run on Chinese-made chips, addressing US export controls that restrict Chinese firms from accessing advanced Western semiconductors. On release day, 10 Chinese chip designers including Cambricon and Biren Technology announced hardware support for U1. The company released the model for free on Hugging Face and GitHub.

Poolside Debuts Laguna Coding Models from US

San Francisco-based startup Poolside launched two new Laguna large language models optimized for agentic coding workflows, marking a notable US entry in the open-source AI space dominated by Chinese companies. VentureBeat reported that the models can write code, use third-party tools, and take autonomous actions beyond simple chat or content generation.

Poolside, founded in 2023, also released a coding agent framework called “pool” and a web-based development environment named “shimmer” that works on mobile devices. The company positions its models as affordable alternatives to proprietary offerings from OpenAI and Anthropic while maintaining competitive performance.

The Laguna models represent a shift in the AI landscape where smaller US companies are adopting open-source strategies previously dominated by Chinese firms. Poolside’s approach targets developers seeking local, cost-effective solutions for coding tasks.

Cisco Addresses AI Model Security Concerns

Cisco on Thursday unveiled its open-source Model Provenance Kit to help organizations track and verify third-party AI models. SecurityWeek reported that the tool addresses security, compliance, and liability issues stemming from unverified claims by model developers and inadequate tracking of model modifications.

The company highlighted risks including poisoned or vulnerable models, training data biases, and licensing complications. Organizations often use models from repositories like HuggingFace without proper documentation of changes or verification of developer claims about sources and vulnerabilities.

“Without provenance, organizations have no easy way to trace an incident back to its root cause, and no way to determine which other models in their stack are also affected,” Cisco explained. The tool aims to improve supply chain integrity and regulatory compliance for AI deployments.

Meta Shifts Strategy with Proprietary Muse Spark

Meta introduced its new AI model Muse Spark at the beginning of Q2, marking a departure from its previous open-source Llama models. CNBC reported that the proprietary approach represents a strategic shift as investors seek clarity on CEO Mark Zuckerberg’s AI monetization plans.

Analysts at Citizens described AI as a “complementary good” for Meta’s core advertising business. The company’s pivot from open-source to proprietary models signals potential revenue generation strategies beyond its traditional social media platforms.

Muse Spark’s performance details remain limited, but the model’s commercial positioning suggests Meta aims to compete directly with OpenAI and Anthropic’s paid offerings rather than supporting the broader open-source ecosystem.

What This Means

This week’s model releases highlight three key trends reshaping AI development. First, Chinese companies continue driving open-source innovation despite US sanctions, with DeepSeek and SenseTime releasing capable models designed for domestic hardware. Second, smaller US firms like Poolside are adopting open-source strategies to compete with tech giants, potentially democratizing access to advanced AI capabilities.

Third, the divide between open and proprietary models is becoming more strategic. While Chinese firms embrace openness partly due to hardware constraints and market positioning, US companies like Meta are reconsidering open-source approaches as investors demand clearer monetization paths.

The technical advances — from DeepSeek’s long-context processing to SenseTime’s direct image reasoning — suggest the performance gap between open and proprietary models continues narrowing. This convergence could accelerate AI adoption across industries while intensifying competition between global AI powers.

FAQ

What makes DeepSeek V4 different from previous models?
V4 can process much longer prompts than earlier versions through a new design optimized for handling large amounts of text efficiently. It represents DeepSeek’s first major release since the breakthrough R1 model in January 2025.

Why is SenseTime’s U1 model significant for image processing?
U1 can generate and interpret images without first converting them to text, which speeds up processing and reduces computing power requirements. This direct image reasoning capability could improve robot perception and physical world understanding.

How does Poolside’s approach differ from other AI companies?
Poolside offers open-source models specifically optimized for agentic coding workflows, targeting developers who want local, affordable alternatives to proprietary services from OpenAI and Anthropic. The company also provides integrated development tools and mobile-optimized environments.