AI's deployment contest is widening beyond closed chat models. Thinking Machines released a customizable multimodal base, Google exposed a long-running research agent, and Emergent raised heavily for end-to-end application building. Fragmented policy and runtime cyber risk complicate that expansion, while world models redirect capital toward systems that represent physical change.
1. Thinking Machines releases Inkling with 975 billion total parameters
Thinking Machines released full weights for Inkling, a multimodal mixture-of-experts model with 975 billion total parameters, 41 billion active per token, and a context window up to one million tokens. The company says it trained the model from scratch on 45 trillion text, image, audio, and video tokens and supports fine-tuning through Tinker.
The Apache 2.0 weights are available through Hugging Face, but the lab explicitly says Inkling is not the strongest open or closed model overall. Most benchmark results are company-run or compiled under differing harnesses, and local deployment at this scale remains demanding. The differentiator is therefore controllability: deployers receive weights, multimodal inputs, adjustable reasoning effort, and a path to customized checkpoints.
Sources: Thinking Machines' Inkling release · Inkling's Hugging Face model card and weights
2. Gemini's Deep Research Agent enters API preview as a background job
Google's Gemini Deep Research Agent is available in preview exclusively through the Interactions API. Its documentation requires developers to set background execution to true, then poll for a result or stream updates, reflecting research tasks that can outlast a normal synchronous model request.
Preview status means interfaces, quotas, and behavior may change before general availability. The execution model is still an important architectural signal: long-running AI research becomes a stateful job with status, cancellation, provenance, and partial-progress requirements. Applications need to distinguish a generated report from the underlying search trail and preserve citations instead of treating one final response as sufficient evidence.
Sources: Google's Gemini Deep Research Agent documentation
3. Emergent raises $130 million for end-to-end AI application building
Emergent raised a $130 million Series C at a $1.5 billion post-money valuation, five times its January valuation, according to TechCrunch. The company says it has more than 200,000 paying customers and a $120 million annualized revenue run rate, and positions its AI platform around building, testing, hosting, debugging, and deploying applications for nontechnical users.
The revenue and customer figures come from Emergent's chief executive and were not independently audited in the report. The round still shows capital moving beyond code completion toward ownership of the full delivery loop. That broader promise also raises the evaluation bar: a generated application must survive security review, maintenance, data migration, and production incidents, not merely produce a convincing first demo.
Sources: TechCrunch on Emergent's Series C and operating claims · Emergent's description of its application platform
4. Anthropic pursues state AI safety laws as OpenAI favors uniformity
Anthropic is pressing for tougher state-level AI safety laws, distinguishing its lobbying strategy from OpenAI's effort to streamline rules nationally, according to Politico. The report establishes a strategic split between two frontier labs, but it does not mean any particular proposal has passed or that state requirements will converge.
The disagreement is both substantive and competitive. State rules can preserve local experimentation and impose stronger duties, while a single federal framework can reduce conflicting compliance regimes. Vendors also enter that debate with different products, risk postures, and market positions, so public-interest claims require comparison against the actual thresholds, audit powers, preemption clauses, and enforcement remedies each company supports.
Sources: Politico on Anthropic's state-by-state AI policy strategy · Anthropic's stated AI policy priorities
5. Carnegie calls for runtime oversight of autonomous cyber agents
A Carnegie Europe paper argues that EU rules built around product certification and human operators do not adequately govern autonomous AI agents acting inside trusted systems. It proposes runtime monitoring, access limits, action logs, shared incident reporting, and procurement requirements, alongside investment in AI-enabled defense and reduced dependence on US frontier models.
The paper synthesizes reported incidents and policy gaps without presenting a new controlled experiment. Its strongest operational point is that predeployment approval cannot capture behavior after an agent gains tools, memory, and changing external inputs. Continuous authorization and suspension mechanisms address that gap left by provider safeguards, especially in critical infrastructure where a trusted agent can cross several systems.
Sources: Carnegie's paper on autonomous cyber operations · Anthropic's documented AI-orchestrated espionage case
6. The world-model debate shifts attention from language to dynamics
A Time Ideas essay argues that world models, systems trained to predict how environments change, are becoming a major direction for AI research and investment. It connects that approach to robotics, weather, biology, and climate, while contrasting it with language models that primarily learn patterns in descriptions of the world.
World model remains a loose and sometimes promotional label, and current systems cannot reliably extrapolate into physical regimes absent from their data. Google DeepMind's Genie 3 illustrates both sides: it generates interactive 720p environments at 24 frames per second for several minutes, but has limited action spaces, imperfect geographic accuracy, and unresolved multi-agent simulation. Data coverage and physical constraints remain as important as scale.
Sources: Time's world-model argument and caveats · Google DeepMind's Genie 3 capabilities and limitations