There is a question every customer asks us at the first demo: "but a model running in this box — how much dumber is it than ChatGPT?" It is the right question, and it deserves numbers with the source attached, not opinions.
The industry's most widely used yardstick is the Artificial Analysis Intelligence Index, which combines several benchmarks into a single score. As of August 2026 (as recorded by BenchLM), the ranking says this:
- The best of the cloud: Claude Opus 5 at 63.0, GPT-5.6 at 58.9, Gemini 3.7 at 56.0. All paid, per user, with your data on their servers.
- The best you can install in an office: DeepSeek V4 Pro at 53.2, GLM-5.2 at 52.6, DeepSeek V4 Flash at 51.8. All open weight, with licences that allow commercial use, runnable on a desktop machine.
Eleven points separate the best of the cloud from the best that runs in an office. In 2024 that gap was an abyss; today it is a margin that, for reading contracts, preparing drafts and querying an archive, you simply do not feel.
How intelligent they are: the cloud versus your own office
Intelligence index — higher is better
Eleven points separate the best of the cloud from the best you can install in your office. For reading contracts, preparing drafts and querying an archive, it is a difference you will not notice — the one you will notice is where your data lives. Source: Artificial Analysis Intelligence Index, as recorded by BenchLM, August 2026. Local inference with ds4, an open-source engine (MIT).
Why it became possible
Two things matured together. The first is model architecture: DeepSeek V4 Flash has 284 billion parameters in total but activates only 13 per answer (Mixture of Experts) — the quality of a giant, the consumption of a mid-sized model. The second is the engine: ds4, the open-source inference engine (MIT licence) by Salvatore Sanfilippo, the creator of Redis, written specifically to run these models on unified-memory machines — with one ingenious choice: when memory runs out, it uses the NVMe disk as an extension, and sustains contexts of up to one million tokens.
On our AG Core Max (128 GB of unified memory) this pairing runs DeepSeek V4 Flash and GLM-5.2 locally, exposing APIs compatible with industry standards: the professional software connects as if it were talking to the cloud. Except that the cloud, this time, sits on your desk.
What the numbers do not say
A score does not capture the variable that matters most to a firm: where the reasoning happens. The 63 points of a cloud model include your documents travelling to third-party servers; the 52 of a local model do not. That is why the right comparison is not "63 versus 52", but "63 with your data away from home versus 52 with your data in the safe". Everyone picks their own yardstick — we have picked ours.