Two releases this week say the same thing from opposite ends of the size chart: open models are now a serious enterprise option.
IBM goes small and sensible
IBM released Granite 4.2 in 3B, 8B and 30B sizes, built for agentic workflows, with native step-by-step "thinking" so the model can plan, weigh trade-offs and catch mistakes before acting [1]. IBM also shipped new speech models at about 470 million parameters, small enough for laptops, phones and edge devices [1]. The models are available on Hugging Face, Ollama and GitHub [1].
IBM's pitch is that it's no longer enough for an AI system to answer clearly. It has to plan, act and finish complex tasks reliably [1]. Agreed. Most business value lives in the boring middle: reading the invoice, checking it against the PO, and routing the exception to a human.
Tencent goes big
At the other end, Tencent open-sourced Hy4 preview, with 770 billion total parameters, 49 billion active, and a context window over a million tokens, aimed at coding, office work and research [2]. That's not something you'll run under your desk, but it's available as open weights and through APIs [2].
Why this matters for normal organizations
- Choice. You can run a small model on your own hardware, a large one through a provider, or both. That's leverage in vendor negotiations.
- Control. Open weights mean you decide where data goes and when versions change.
- Responsibility. Also means you own patching, monitoring and evaluation. "Open" is free like a puppy, not free like a sandwich.
A fair caution: where a model comes from matters to some organizations for legal, policy or risk reasons. Know your requirements before you download anything, and evaluate on your own tasks, not a leaderboard.
My rule of thumb this year: start with the smallest model that passes your own test set. Scale up only when the results make you, not the vendor deck.
Where I'd start
Pick one repetitive, document-heavy process with clear right answers, like invoice matching or ticket triage. Run a small open model on it in a sandbox with real (sanitized) examples for two weeks. Measure accuracy, speed and how often a human had to step in. If the numbers hold, you have a business case. If not, you spent two weeks instead of two quarters.
And keep a human on the exceptions. That's where the judgment lives.
The questions to ask before you adopt one
Who maintains it, how often it's updated, what license it carries, and who on your team owns patching it. An open model you download and forget is just another unpatched server with better marketing. Treat it like any other software dependency, and it can save real money.