SPEAKER_00 0:00

Welcome back to Sumo Growth AI Insights. I'm Darius, and I'm walking you through this week's AI news the way an operator actually uses it, not as a pile of headlines. Lead Story Meta releases Muse Spark Open, the most capable US OpenWaight model to date. On Thursday, Meta released the weights for Muse Spark Open, calling it the most capable US OpenWait model to date, built for coding and long-running agent work. The model has 816 billion total parameters, with 42 billion active per token, a 1 million token context window, and native text, image, audio, and video input. Meta says it's the first model in its class that can run on a desktop using two Apple M5 Ultra Max Studios. The weights are available on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden, and Microsoft Foundry, with hosted APIs from providers including Grok, Fireworks AI, and OpenRouter. For agencies, this is a serious open option to pilot on one internal coding or research agent before you put any client data near it. Plus story Microsoft finishes MAI Base 2, a 2.8 trillion parameter in-house base model. On Thursday, Microsoft AI said it has completed pre-training MAI Base 2, a mixture of experts foundation model with 2.8 trillion total parameters and 80 billion active, trained on 16,000 NVIDIA GB200 GPUs over 38 days. Microsoft says the model beats every open weight model it tested on held out evaluations and that it's 4 to 12 times more pre-training efficient than DeepSeek V4 Pro Base. Microsoft plans to license the weights to trusted partners through Foundry rather than open sourcing them. The model is still in post-training, so there's no public API yet. For agencies, Microsoft is building deeper in-house intelligence for Copilot and Azure, so expect more Microsoft-only model options in the stacks your clients already pay for. Three more headlines. First, Anthropic is giving METR researchers employee-level access to audit its models. On Thursday, Anthropic said it's embedding six researchers from the nonprofit METR inside the company with employee-level tools and office access, and it won't veto their published assessments of Anthropic's risks and practices. Second, the FTC subpoenaed OpenAI and Anthropic over AI consumer safety harms. On Wednesday, the FTC sent 50-page civil investigative demands seeking records on security incidents, safeguards, marketing, and privacy policies, and it plans to compel AI executives to testify. Third, NVIDIA launched the Open Nematron 3.5 Super VL for multimodal agents. It's a 120 billion parameter open multimodal model for documents, charts, and tool use, plus smaller bolt variants for local agents, released under NVIDIA's OpenMDW license. Tools of the day. Jockey from 12 Labs at 12labs.io lets you search, analyze, and generate text from your entire video library with multimodal AI. Harness at autonomous.ai slash harness lets you run every coding agent across every machine from one open source command center. Pomerado at Pomerado.ai turns any website into an API or MCP tool your agents can call reliably. Business AI tip of the day run a one-week open model pilot before you promise clients. Meta just put Muse Spark OpenWaights in developers' hands. The menu of run-it-yourself models is growing fast, and clients will ask whether your agents can leave the big closed APIs. So treat the next week as a controlled pilot, not a migration. Pick one internal workflow only, like lead research, FAQ drafting, or video library search with zero live client personal data. Run the same 10 prompts on your current model and on one open option and score accuracy, latency, and cost side by side. Write a kill switch. Which human approves outputs, what data the model can see, and how you shut the agent off in one click. Log every failure mode in Sumo AI CRM or GoHi level so you can show the client what broke and what you fixed. Decide go or no go in writing before you mention the model on a sales call or in a proposal. And if you use a tool like Harness or Pomerado in the pilot, keep credentials in a vault and rotate them when the test ends. Open weights are a leverage play, not a free lunch. The agencies that score the pilot first will sell the upgrade with proof, not hype. Pro tip Put the scorecard in the discovery deck. Prospects care less about parameter counts than about a screenshot of your model beating the old stack on their exact task list. That's your Sumo Growth AI Insights briefing. Run one honest pilot this week, and I'll catch you on the next issue.