Open Models Surge as OpenAI Sunsets Agent Builder

<p>Google's <strong>Gemma 4 12B</strong> lands as a free, encoder-free multimodal model (text, image, audio, video) on 16 GB VRAM under Apache 2.0, while Cohere's <strong>Command A+</strong> becomes the first fully open enterprise-grade large model — giving builders two credible self-hosted paths in one week. Meanwhile, OpenAI is sunsetting Agent Builder ...

AI Industry Dossier

Highlights

  • Google releases Gemma 4 12B under Apache 2.0: an encoder-free model that natively processes text, image, audio, and video in a single backbone, running on 16 GB VRAM with a 256K context window. (Google Blog)
  • OpenAI is deprecating Agent Builder and Evals on November 30, 2026, directing no-code builders to Workspace Agents and code-first teams to the Agents SDK. (OpenAI)
  • Cohere releases Command A+ as its first fully Apache 2.0-licensed open enterprise model — 218B sparse MoE, 25B active parameters, native vision and reasoning, designed to fit on 2× H100s. (VentureBeat)
  • Snowflake acquires Natoma to embed a centralized MCP gateway — with tool-call-level identity, policy, and audit — directly in its Cortex Agents data platform. (Snowflake)

Key Signals

  1. June 3, 2026
    Gemma 4 12B: encoder-free open multimodal from Google

    Unlike prior architectures that route images and audio through separate encoders, Gemma 4 12B handles all modalities directly in the LLM backbone — simplifying inference pipelines and enabling native audio input that no prior mid-sized Google open model has offered. Apache 2.0 licensing and weights on Hugging Face and Kaggle make it an immediate candidate for sovereign, air-gapped, or cost-sensitive deployments. (Google Blog)

  2. June 3, 2026 notice · Nov 30 deadline
    OpenAI sunsets Agent Builder and Evals

    Teams running production workflows on OpenAI's hosted Agent Builder or Evals products have five months to migrate. OpenAI is routing low-code users to the Workspace Agents surface — Codex-powered, with Slack and Salesforce integration, free period extended to July 6 — and directing code-first teams to the Agents SDK. This is a product consolidation, not a capability removal, but the migration window is active now. (OpenAI)

  3. Announced May 27 · Snowflake Summit, June 2–3
    Snowflake + Natoma: MCP governance lands in the enterprise data stack

    Natoma's centralized MCP gateway enforces identity, policy, and audit at the tool-call level across SaaS apps, cloud VPCs, and on-premises systems. Paired with AI Agent Identity — per-agent RBAC with cryptographic audit trails — unveiled at Summit, Snowflake is threading governance into every agentic action that touches its platform, a significant step beyond ad hoc tool-call logging. (Snowflake)

Why It Matters / What To Watch

  1. Open enterprise models now cover both edge and large-scale deployment
    • Gemma 4 12B (16 GB VRAM, encoder-free multimodal) and Cohere Command A+ (2× H100s, 48 languages, full Apache 2.0) give operators two credible open paths unavailable six months ago — teams locked into API-only contracts for regulatory or cost reasons should evaluate both. (Google Blog) (VentureBeat)
    • vLLM v0.22.0 (released May 29) ships an experimental Rust frontend, DeepSeek V4 production hardening, and XGrammar 0.2.0 for strict structured output — the open serving layer is tracking model-quality improvements closely. (GitHub)
  2. The agentic product surface is consolidating; governance is becoming the moat
    • The OpenAI Agent Builder/Evals sunset follows a visible pattern: feature-sparse hosted builders are being replaced by SDK-first runtimes and governed enterprise surfaces. Audit current Agent Builder usage and begin Agents SDK migration planning now — five months is shorter than it sounds for production systems. (OpenAI)
    • Groq is raising $650M to pivot from chip vendor to managed inference cloud following NVIDIA's $20B technology licensing deal. When a purpose-built inference hardware company repositions around managed service, it signals that reliability, governance, and SLA coverage are becoming the competitive layer over raw throughput. (The AI Insider)

Quick Links