Open weights go local, Google halves agent costs, Databricks reaches $190B, and Anthropic goes shopping for throughput.

The AI race just moved from the cloud to the device in your bag. Meta released a 30-billion-parameter agentic model that can run locally under an Apache 2.0 license, while Google cut the introductory price of its newest coding-and-agent model in half versus its previous Flash release. At the same time, Databricks raised $5 billion at a $190 billion valuation to own the data layer beneath enterprise AI, Anthropic entered talks to buy an infrastructure specialist, DeepSeek pushed its newest Pro model into general availability, and California gave AI cyber defense its own operating program. The thread connecting it all: AI is no longer a single model decision. It is becoming an architecture decision — where the model runs, what it can access, how much each workflow costs, and who controls the data around it.

Meta Superintelligence Labs released Muse Glimmer, a 30-billion-parameter, open-weight model designed for always-on local agents. The weights are available under Apache 2.0, and Meta says the model is optimized for local coding, tool use, function calling, multimodal input, and long-horizon workflows. The company says quantization cuts the language model to under 20 GB, though it recommends a 24 GB or 32 GB memory envelope once working memory and its perception components are accounted for. In practical terms, that brings a capable agent closer to a workstation, laptop, or on-premises deployment than a permanently metered cloud endpoint.

This is a strategic reversal with teeth. The highest-value enterprise workloads increasingly need models that can see internal files, operate tools, and retain context. Running those workloads locally can reduce inference cost and data exposure, but it shifts the burden to IT: hardware selection, patching, access control, monitoring, and model governance become the buyer's problem. Meta is betting that enterprises and developers will accept that trade in return for deployability and control — and that American open weights can compete against the Chinese models already winning on price and accessibility.

The business question is no longer open versus closed as an ideology. It is which workloads deserve a local model, which require frontier-cloud capability, and how to keep both inside one governance model. Muse Glimmer will not replace a top cloud model for every task. It does make a serious case for moving routine, data-sensitive, or persistent agent work nearer to the business that owns it.

Read the full story →

The parameter count of Meta's Muse Glimmer — an open-weight agentic model designed to bring local AI agents to consumer-grade hardware rather than keep every workflow inside a cloud API.

Source: Meta AI

Google released Gemini 3.7 Flash — its newest “workhorse” model for coding and agents — just three weeks after 3.6 Flash. The launch price is $0.75 per million input tokens and $3.75 per million output tokens through year-end, half the original 3.6 Flash price, while Google reports gains across coding, web development, document understanding, and workflow-automation benchmarks. Those benchmark claims are Google's own, but the price move is indisputable: agent platforms are now competing on cost per completed workflow, not merely model prestige.

Databricks closed a $5 billion funding round at a $190 billion valuation, up from roughly $134 billion six months ago. The company said it surpassed a $7 billion annualized revenue run rate and more than 80% year-over-year growth in its second quarter; Reuters reported the round was led by existing investors Coatue, Blackstone, MGX and accounts advised by T. Rowe Price, plus new investor Sixth Street Growth. The investor message is clear: the enterprises that control data, governance, and the agent control plane are becoming just as strategically valuable as the labs that train the models.

Anthropic is in early talks to acquire Nvidia-backed Decart AI, Reuters reported, citing a source familiar with the matter; Bloomberg first reported that a deal could be worth about $6 billion. Decart works on AI infrastructure and optimization, live video editing through Lucy, and simulated environments for robotics and autonomous-driving development through Oasis. Nothing is signed and the talks could still collapse, but the logic is revealing: frontier labs are treating inference performance, video, and simulation capacity as strategic assets rather than commodity infrastructure.

DeepSeek formally released V4 Pro to its app, web client, and API today, with new pricing scheduled to take effect on August 16. The company is moving quickly to keep pace with domestic rivals and frontier U.S. labs — a reminder that AI procurement cannot be a one-time vendor decision when new capability and price points arrive this fast. For cost-sensitive technical teams, compare the real workload economics before defaulting to any single provider.

California announced an AI Cyber Defense Program inside the California Cybersecurity Integration Center, aimed at vulnerability detection, network hardening, and protection of critical infrastructure. The move follows a run of public concern about AI-enabled cyber risk and shifts the conversation from policy principles to operational capability. Security leaders should read this as a signal: AI defense is becoming a standing function, not an experimental initiative.

This week's toolbox is for teams moving from “let's test an agent” to “how do we run agents without losing control?” We have one open-source control plane for skills and permissions, one agent-first workspace for developers, and one governed enterprise layer for business teams. Different entry points, same job: make agents useful without making the organization less secure.

Last week’s headlines made it sound like three frontier labs had watched their models independently break out of containment. CNBC's reporting adds the missing context: OpenAI, Anthropic, and Meta all referenced the same evaluation vendor, Irregular, which said the cases stemmed from the same evaluation-environment issue and did not constitute a sandbox escape or sophisticated cyber action. That does not make the incidents trivial — the models still exposed gaps in the testing environment — but it is the distinction every executive needs before turning a safety headline into a board-level conclusion.