AI Strategy
Open-Weight AI Just Got Good Enough to Matter for Your Firm
A busy month for models you can download
For most of the last few years, the most capable AI models were only available one way: you sent your data to a vendor's API and got an answer back. Open models existed, but they trailed the closed ones by enough that most firms didn't seriously consider them for client work.
October 2026 made that gap a lot harder to ignore. In a single month, Mistral published the weights for Mistral Large 4, its trillion-parameter flagship. Meta released Muse Spark Open. DeepSeek shipped another update to its V4.1 Flash line under an MIT license. NVIDIA released Nemotron 3.5 Bolt, a model small enough to run on a single 24GB graphics card. Reflection, a startup founded by former Google DeepMind researchers, debuted Beam and said the weights would follow.
The vendors' benchmark claims are their own and should be read that way. But the direction is clear, and enterprises are acting on it. A recent ETR survey of 200 enterprises found open-weight models now account for about a third of their AI token usage, up from under a quarter.
"Open-weight" is not the same as "open source"
The term gets used loosely, so it's worth being precise. An open-weight model is one where the company publishes the trained model itself, so you can download it and run it on hardware you choose. That is different from open source in the traditional sense, which would also mean publishing the training data and the code used to build it. Almost none of the major releases do that.
For a firm deciding whether to use one, the practical questions are simpler: Can we run it where we want? Can we use it commercially? Can anyone take it away from us?
Read the license before you read the benchmarks
This is where firms get caught. "Open" covers a wide range of licenses:
- Permissive licenses like Apache 2.0 and MIT let you use, modify and deploy the model commercially with very few strings attached. DeepSeek's V4.1 Flash and Aleph Alpha's Kolibri-1 fall in this group.
- Custom licenses from Meta, Moonshot, Z.ai and others are usually fine for a firm using the model internally, but they often add attribution requirements or extra terms above certain revenue thresholds, particularly if you resell the model as a service.
For a 40-person accounting or law firm running a model for its own workflows, most of these licenses are workable. The point is to check before you build, not after a client's security questionnaire asks.
Why this matters for professional services
We've written before about who actually owns your client data. Open-weight models change that conversation in three ways.
Control over where data goes. A model running in your own cloud account, or on a server in your office, never sends client documents to a third party. For firms handling tax returns, privileged legal material or health information, that's often the difference between "we can't use AI for this" and "we can."
No one can switch it off. Mistral's VP of Science put it bluntly when announcing Large 4: access "can't be cut off by definition because the model is open-weighted." If a vendor changes its pricing, retires a model you depend on, or changes its terms, the copy you downloaded keeps working.
Predictable costs. API pricing is per use. Self-hosted models have a fixed infrastructure cost. For high-volume, repetitive work like document classification or data extraction, the math can tilt toward running your own.
What it doesn't change
Open-weight models don't remove the hard parts of AI adoption. Someone still has to host, monitor, patch and update the system. The largest models need serious hardware: Meta's Muse Spark Open needs roughly 700GB of GPU memory to run. And the frontier closed models are still the best choice for some of the most demanding reasoning tasks.
The right answer for most firms isn't "go all open" or "stay all closed." It's matching each workflow to the right model: a small, self-hosted model for sensitive, repetitive tasks, and a frontier API for complex work where the data can be appropriately handled.
What's new is that this is now a real decision, not a theoretical one. We'll walk through how to make it in a practical guide next month.