Open models are arriving at laptop scale with context windows that match cloud-only offerings from a year ago; the world's largest messaging platform is activating as an autonomous commerce layer; and data-platform agents are collapsing the distance between a warehouse and a workflow. Four developments from the last 48 hours that change the calculus.
Gemma 4 12B: frontier multimodal at laptop scale, Apache 2.0
Google shipped Gemma 4 12B on 3 June under an Apache 2.0 licence, a notable step beyond the custom Gemma licence used in prior releases, which carried commercial restrictions. The model handles text, images, audio, and video through a single encoder-free backbone; unlike its predecessors, it requires no separate vision or audio encoder, which reduces multimodal latency and simplifies deployment. It is the first mid-sized model in the Gemma family with native audio input.
At 11.95 billion parameters, the model fits on a standard enterprise laptop with 16 GB of memory and benchmarks close to Google's own 26B mixture-of-experts Gemma 4 variant. The context window is 256,000 tokens, enough for an hourlong meeting transcript, a mid-sized codebase, or a 200-page regulatory filing in a single pass. Weights are available on Hugging Face and Kaggle; the model integrates with vLLM, SGLang, MLX, and llama.cpp.
For organisations that cannot route data through external APIs, a locally deployable multimodal model under Apache 2.0 removes both the licensing friction and the latency argument against open models. This is the most commercially clean laptop-scale model available today.
NVIDIA Nemotron 3 Ultra: self-hosted frontier agents, available now
NVIDIA Nemotron 3 Ultra became generally available on 4 June via Hugging Face and NVIDIA NIM microservices, two weeks after Jensen Huang previewed it at Computex 2026 in Taipei. The model carries 550 billion parameters in a mixture-of-experts architecture, with 55 billion active per inference. Open weights, training data, and recipes are published together on Hugging Face.
NVIDIA cites 5x faster inference throughput and 30% lower cost per complex agentic task versus the prior Nemotron generation. The model targets long-running agents across coding, research synthesis, and enterprise workflows, where latency across a multi-step task sequence compounds into real cost. Platform integrations include DataRobot, H2O.ai, UiPath, and a broad set of NVIDIA Cloud Partners.
Running a 550B mixture-of-experts model at NIM throughput on your own infrastructure is no longer a research project. For enterprises carrying data-residency or compliance commitments, this shifts the cost-capability trade-off at the frontier end of the open-model stack in a way that hosted-only options could not.
Meta Business Agent goes global: one million businesses, two billion users
Meta announced the global rollout of Meta Business Agent on 3 June at its Conversations conference in London. The agent is now available to businesses of all sizes on WhatsApp, Instagram, and Messenger, the platform combination that collectively handles more consumer messaging than any other stack worldwide.
The agent handles inbound inquiries, surfaces products from a business catalogue, books appointments, vets sales leads, completes transactions, and escalates to a human operator at a threshold the business defines. Integrations with Shopify, Zendesk, and Shopee are available through the Meta Business Agent Platform. More than one million businesses were already running the predecessor chatbot on WhatsApp and Messenger before this global rollout. The service is free at launch; paid tiers are coming.
The operator implication is precise: inbound customer service and commerce flows that previously required a dedicated support team or a custom deployment now run on infrastructure that a small or mid-market business can activate within minutes on a channel its customers already use daily. At scale, that is a margin story, not just an efficiency story.
Snowflake Summit 26: CoCo and CoWork collapse the gap between data and agent
At Snowflake Summit 26 on 2 June, Snowflake released CoCo (formerly Cortex Code), a coding agent embedded in the Snowflake platform, alongside CoWork (formerly Snowflake Intelligence), a personal agent for knowledge workers. Both draw on Cortex Sense, a new shared context layer that automatically surfaces an organisation's data definitions, business rules, and operational knowledge, anchoring agent outputs to the enterprise's own data rather than to generic world knowledge.
CoCo lets builders automate workflows, develop applications, and operationalise AI on enterprise data through a prompt, without leaving the Snowflake environment. CoWork moves business teams from data query to decision to action in a single interface. Sanofi announced on the same day that it is deploying Snowflake to accelerate drug-discovery workflows, a signal of where data-native agents are landing first.
For enterprises already on Snowflake, the compound effect is concrete: the data platform and the agent layer are the same system, reducing the number of integration points and security perimeters a team must manage.
The through-line is convergence at the edge and the platform layer. Open models are matching cloud-only capability at laptop scale and self-hosted cost. Platform agents are embedding themselves in the distribution channels and data systems enterprises already operate. The build-vs-buy question has not simplified, but it has sharpened: the meaningful choice is now which platform layer you want owning your agent context.