NeuralStackA cross-section of the latest AI model releases shows how quickly the field is tilting toward large, specialized systems with increasingly open access to their weights. ByteDance’s Seed lab is pushing video generation further with Seedance 2.5, which now extends single-take clips to roughly 30 seconds while referencing a wide set of multimodal assets. MiniMax is expanding into open-weights video with MiniMax H3, promising 2K clips, native stereo audio, and support for editing and motion transfer in segments up to 15 seconds. In parallel, DeepSeek has graduated its V4-Flash model from preview to an official release, with a new checkpoint tuned specifically for agent-style use cases.
Open-weight large language models are at the center of this wave. Thinking Machines’ Inkling-Small is a 276B mixture-of-experts model with 12B active parameters that ships full Apache-2.0 weights, yet is positioned to match the performance of a sibling model roughly four times its size. Moonshot AI has moved the largest open-weight model in history into the wild: Kimi K3, a massive 2.8T-parameter system now available for self-hosting via public checkpoints. Upstage is contributing Solar Open 2, a 250B/15B open-weight MoE with a million-token context window and a hybrid attention design aimed squarely at agentic work. Poolside’s Laguna S 2.1, a 118B MoE with 8B active parameters and a similarly long context, is pitched as a Western coding counterpart to the agentic stacks associated with DeepSeek.
That open-weight push is not limited to giant models. Nanbeige’s Nanbeige4.2-3B uses a looped-transformer architecture to recycle layers and hit competitive scores on benchmarks like SWE-Bench Verified despite being far smaller than many of its rivals. Cisco’s Antares models, at 350M and 1B parameters, are designed to run on-premises and walk real code repositories to locate files that may conceal vulnerabilities flagged by common weakness enumerations. Kwaipilot’s KAT-Coder-V2.5-Dev adds another open-weight coding MoE at 35B parameters and 3B active, tuned to operate inside production codebases rather than only autocomplete short snippets. And Inflect-Micro-v2 shows how far compression can go, packing a complete English text-to-speech system into roughly 37.5 MB while still beating larger systems in most blind human tests.
Robotics is another area seeing coordinated model launches. DeepMind’s Gemini Robotics ER 2 is described as an embodied-reasoning "planning brain" that can watch video, plan multi-step work, and orchestrate multiple robots simultaneously. Gemini Robotics 2, a humanoid-focused model, extends that control to full-body motion, including walking, crouching, tying knots, and coordinating two robots at once. Black Forest Labs is pushing its FLUX 3 backbone beyond pure generation: the core FLUX 3 model spans video with audio, image editing, and real-world robotic control at an industrial site, while FLUX 3 x mimic focuses on driving a factory arm from a single on-robot GPU in near-real time. These releases together suggest that robot orchestration is becoming a first-class target for general-purpose multimodal models.
Specialized models are also reshaping adjacent domains like security, imaging, and media. Microsoft’s MAI-Cyber-1-Flash targets cybersecurity explicitly, integrating with the MDASH stack to find software bugs more cheaply and effectively than previous GPT-based setups. A separate Microsoft model, Mage-VL, treats video as a codec stream, preserving anchor frames while discarding many predicted-frame patches to cut visual tokens by around three quarters, a shift that could reshape how multimodal systems handle long-form video. DeepHealth, part of RadNet, has secured FDA clearance for its breast ultrasound AI, which automates lesion reads and is being rolled out across the company’s imaging centers. Meanwhile, Google DeepMind’s Lyria 3.5 brings richer melodies, clearer vocals and longer compositions to the Flow Music app, and Grok Voice Think Fast 2.0 from xAI moves speech-to-speech latency into sub-second territory while topping a specialist leaderboard.
Commercial models are not standing still on pricing and access. OpenAI has cut prices on its GPT-5.6 Luna tier by around 80% and Terra by roughly 20%, only three weeks after launch, citing about 20% gains in serving costs and passing those savings through the API. Those reductions make it cheaper to run high-end proprietary models at scale, and likely reflect competitive pressure from both rival closed systems and the growing pool of capable open-weight alternatives. At the same time, Switzerland’s Apertus 1.5 shows how national initiatives can push fully open models into multimodality while preserving permissive Apache-2.0 licensing. Together, these moves underscore that price, openness, and performance are all active levers in a shifting market.
Why this matters
The burst of releases across frontier-scale open weights, compact agents, and domain-specific models signals a more pluralistic AI ecosystem where organizations can mix and match building blocks rather than rely on a single vendor. Labs like Moonshot, Thinking Machines, Upstage, Nanbeige and Cisco are effectively turning capabilities that once required proprietary stacks into downloadable infrastructure that can be self-hosted, audited, and tuned in-house. When those open-weight systems are paired with increasingly capable robotics, video and security models from giants like Google, Microsoft and DeepSeek, the result is a toolkit that lets enterprises build agentic workflows spanning code repositories, industrial robots, media pipelines and regulated healthcare. The competitive response from proprietary providers on pricing suggests that open weights are not just a philosophical choice but a practical force reshaping margins and deployment strategies.
Looking ahead, this mix of giant MoEs, small on-device models, and multimodal robotics backbones points toward more integrated, agentic applications that can see, plan and act across software and the physical world. As more of these models release under permissive licenses, expect enterprises and national initiatives to experiment with custom stacks that blend local security-focused agents, long-context coding systems, and factory-ready controllers like FLUX 3 x mimic. Medical tools such as DeepHealth’s cleared breast ultrasound reader hint at a regulated path for clinical AI, while low-latency voice models and creative systems like Lyria 3.5 suggest consumer experiences will keep evolving in parallel. The pace and diversity of these launches make clear that "open-weight vs. closed" is no longer a simple binary, but a spectrum of trade-offs that AI builders will navigate with increasing nuance.