Eleven points: how far open models really are from the cloud
The question we get in every demo is whether the intelligence that fits on a desktop machine is “the real one”. There is a number that answers it, and it is worth looking at whole: what it says, and what it does not say.
The number
The Artificial Analysis Intelligence Index is a composite index that aggregates several reasoning, coding and knowledge tests into a single scale. In the August 2026 measurement, the cloud models are at the top and the ones you can install in an office are below:
| Model | Where it runs | Index |
|---|---|---|
| Claude Opus 5 | cloud | 63.0 |
| GPT-5.6 | cloud | 58.9 |
| Gemini 3.7 | cloud | 56.0 |
| DeepSeek V4 Pro | Conectao Core Enterprise | 53.2 |
| GLM-5.2 | open weights — outside our range | 52.6 |
| DeepSeek V4 Flash | Conectao Core Max | 51.8 |
Source: Artificial Analysis Intelligence Index, BenchLM measurement, August 2026. Read on 30/08/2026.
On the GLM line. GLM-5.2 (June 2026, MIT licence) has open weights, but its lightest file weighs 196.6 GiB and does not fit in the 128 GB of a Core Max: we do not attribute it to our machines. The current line of the family is GLM-5.3-Flash, with weights published on 25/08/2026 and also under an MIT licence. It does not yet have an independent measurement in this index: that is why it is not in the table, and why we do not lend it another model's number. When it has one, we add it with its date.
Between the best of the cloud and the best you can install there are 11 points (63.0 against 53.2). Two years ago that distance was much larger and the conversation made no sense. Today it does, and that is why this business exists.
What eleven points mean in real work
A composite index does not translate into “11% worse”. It means that on the hardest tasks in the set — long multi-step reasoning, non-trivial coding problems, trick questions — the cloud model gets it right more often.
On the tasks a company actually buys these systems for, the difference is hard to notice: reading a contract and flagging the risk clauses, answering a guest on WhatsApp with the hotel's information, extracting the data from an invoice, summarising minutes, drafting a standard document, searching the company archive. These are tasks of comprehension and formatting, not reasoning at the limit.
Put the other way round, and honestly: if your use case lives inside that margin of eleven points, a machine in the office is not the right answer, and we will tell you so in the demo. The way to find out is not this article — it is to test your real case against a machine that is switched on.
What size of model each machine carries
The “size” of a model is measured in parameters: more parameters, more capability. The limit of what fits on a machine is set by memory.
| Configuration | Models it can load | In practice |
|---|---|---|
| Core One | 7–13 billion | Front-desk service, one document at a time |
| Core Studio | 13–27 billion | Contracts, drafts, personal archive |
| Core Team | 32–70 billion | Full company archive, several users |
| Core Max | large MoE models, in quantised format | DeepSeek V4-Flash or GLM-5.3-Flash in Q2 — the blue row in the table above |
| Core Enterprise | beyond that, spread over several machines | Cluster, several teams |
“MoE” means mixture of experts: the model has hundreds of billions of parameters — 304 billion for DeepSeek V4-Flash — but activates only part of them for each answer. That, together with quantisation (the weight file is reduced to Q2 format), is the technical reason a model of that size fits today on a desktop machine and not in a datacentre. Said precisely: the full uncompressed file does not fit in 128 GB — what gets installed is the quantised version.
Parameter counts read at the source (HuggingFace model cards
deepseek-ai/DeepSeek-V4-Flash) on 02/09/2026.
The engine that runs them
Models do not run on their own: an inference engine is needed. We use ds4, open source under an MIT licence, written for one thing: running large models on an office machine. It exposes APIs compatible with the OpenAI and Anthropic standards — which means the software connects as if it were talking to the cloud, without anything leaving the building — and it supports contexts of up to one million tokens, that is, hundreds of pages in a single question.
When memory runs out, the engine uses the NVMe disk as an extension. That allows models larger than RAM to be loaded, in exchange for speed.
Why “open source” is not an ideological detail
An open model is downloaded and it stays. It cannot be withdrawn, made more expensive, or changed underneath you. If tomorrow the provider of a cloud model modifies the price or the terms — and in 2026 all four of the big ones did, as we set out in the other article — whoever has the model on their own machine never notices. The same applies to the software: we install public, verifiable open-source platforms, and that is why there is no per-user licence to pay.
When a better generation of models comes out, the update arrives with the maintenance. Without touching your data, and without renegotiating anything.