NeuralStackA dense run of AI model releases is reshaping the market around open weights, multimodal capabilities and agentic workloads. AI/TLDR’s roundup tracks 324 releases and highlights new systems from NVIDIA, DeepSeek, Meta, Mistral AI, Google, Anthropic, xAI and other companies.
NVIDIA’s Nemotron 3.5 Super VL is an open vision model designed to read text, documents, charts and short videos in one checkpoint, with 12B parameters active per token. DeepSeek’s V4.1 Flash 1017 is an in-place upgrade that raises agent and coding scores while producing shorter outputs at the same price.
Meta is pushing open weights at a far larger scale with Muse Spark Open, an 816B model built for coding and agents with a 1M-token context. The company also introduced Muse Wonder, its first frontier model, aimed at long agent tasks that require remembering relevant information and continuing work until completion.
Mistral AI’s releases cover both frontier-scale and highly specialized use cases. Mistral Large 4 is a 1T multimodal mixture-of-experts model with 49B active parameters and a 1M-token context, while the smaller Timestral family is designed for zero-shot time-series forecasting and can run on a CPU.
Other open-weight releases are targeting practical deployment constraints. Liquid AI’s LFM2.5-24B-A2B activates about 2B parameters per token for laptop-based agents, OpenBMB’s MiniCPM-V 4.7 brings a 35B mixture-of-experts vision model to an iPhone’s memory budget, and Black Forest Labs’ FLUX 3 Image Klein is designed for image creation, editing and combination on a consumer GPU.
Commercial providers are also broadening access to specialized systems. Google’s Gemini 4 Argon reached general availability across the Gemini API, AI Studio, Antigravity and the Gemini app, while Anthropic’s Claude Haiku 5.5 adds a 1M-token context window plus computer and browser use. xAI’s Grok Voice 3.0, meanwhile, is positioned as a leading speech-to-speech model at its price.
Why this matters
The release pattern points to a market moving beyond the single race for the largest general-purpose chatbot. Decision models from Microsoft, Cloudflare and AWS are built to handle routing, grading, guardrails and other choices inside agent workflows, while models from Fish Audio, Kandinsky Lab and Krea focus on speech, video and image production.
Efficiency is becoming as important as scale. Sparse architectures, smaller active parameter counts and models designed for phones, laptops or consumer GPUs make advanced capabilities available in more environments, while open weights give developers greater control over deployment and customization.
The next phase will likely be defined by how reliably these models operate inside products rather than by release size alone. With long-context systems, multimodal agents and specialized decision models arriving in quick succession, developers will have more options to assemble AI systems tailored to specific tasks and hardware.