Private AI for physics · Open models

The open-model staircase for private scientific work

For a fluid-dynamics team that will not send its code, equations or results to an AI lab: four open-weight models, each a step up in capability and hardware, and what it costs to run each one on a node you control.

October 2026Benchmarks: Artificial Analysis, snapshot 1 Oct 2026Sizes: Hugging Face file sizes

The commercial AI APIs ask you to trust that your prompts are not logged, kept or used for training. For a startup whose value is its solvers and its results, that is a trust problem, not a technical one. Open-weight models remove it: you download the weights and run them on a machine you control, so nothing you type ever reaches the lab that made the model.

There are two ways to do that, and this page prices both:

Rent a dedicated node from a cloud provider, run the model server on it yourself, and switch request logging off. You pay by the hour and can stop at any time.
Own the hardware. You pay once, and nothing leaves your building.

Every size below is for 4-bit precision or higher, the lowest we would trust for mathematics. Parameters are given as “total / active”: total is what has to fit in GPU memory, active is what runs for each token and sets the speed.

Artificial Analysis Intelligence Index by step

Bubble size is total parameters. The green line is the best closed frontier model, for reference. Hover any point.

The staircase Best closed model (not an option here, shown for scale)

Why the top step is a big model

Kimi K3 is not the top scorer on every benchmark; some much smaller models come close or beat it on individual tests. We still put it at the top of the staircase, because benchmark scores and research usefulness are not the same thing.

Smaller models are increasingly tuned on the very kinds of problems used to rank them. They score well on what is measured and fall away on what is not. A model with 2.8 trillion parameters holds far more of the literature and more ways of attacking a problem. On Artificial Analysis’s knowledge test, Kimi K3 answers 48% of questions correctly against 35% for MiMo-V2.6-Pro, the 1-trillion-parameter model that tops the open-model Intelligence Index. In our experience that difference shows up exactly where a research team needs it: a result nobody expected, or code for a method that has not been written before. A model that has seen more is less likely to bend an unfamiliar problem into a familiar one.

What about MiMo-V2.6-Pro? Xiaomi’s 1.02T-parameter model (42B active) fits on one 8-GPU RTX PRO 6000 server at 574 GB, and it tops the open-model Intelligence Index (46.3 vs Kimi K3’s 43.6). We do not recommend it for this work. It knows considerably less: 35% correct answers on AA-Omniscience against Kimi K3’s 48%, and an overall Omniscience score of 8 against 20. In our hands-on use it has also been less dependable on open-ended coding than its scores suggest. Benchmark-leading and research-ready are different things.

What each step costs

Steps 1–3 run on NVIDIA RTX PRO 6000 Blackwell cards (96 GB each), the largest GPU that goes in an ordinary server, at 1, 2 and 4 cards. Step 4 does not fit on any RTX PRO 6000 server, so it moves to an 8× B300 node, the standard 8-GPU data-centre system with 2.3 TB of GPU memory. Each model uses the highest-precision release that fits with at least 20% spare memory for context.

Step	Release	Weights	Hardware	Spare for context	Rent / hour	Rent / 8 h day	Rent / 24 h	Own (hardware)	Break-even 24/7

Rent: RTX PRO 6000 at $2.75 per GPU-hour (Hugging Face Inference Endpoints, dedicated instance). B300 at $7.85 per GPU-hour (Nebius on-demand); on-demand prices for a dedicated 8× B300 node range from $57 an hour (Modal) to $142 (AWS). Own: street prices, September 2026. An RTX PRO 6000 is about $16k plus a $10–30k chassis; an 8× B300 HGX system is about $550–650k. Break-even includes electricity at $0.25/kWh with 35% for cooling.

Rent a dedicated node, or own it? Cumulative cost over two years

Owning includes power, running 24/7. Change any assumption in the cost calculator.

Rent 24/7 Rent 8 h × 22 workdays Own: hardware + power

Renting a node around the clock costs as much as buying it, power included, within about 11 to 14 months at every step. A team that keeps a model running all day should plan to own it. Renting only during working hours takes three to four years to reach the purchase price, so renting is the right way to try a step before committing to it.

The jump to step 4 is the expensive one: from about $94k to about $600k to own, and from $11 to $63 an hour to rent. That price buys the model with the most knowledge you can run privately.

Keeping it private

Running an open model removes the AI lab from the picture, but the machine still has to be set up so nothing leaks.

On a rented node, rent bare GPU machines and run the model server yourself, rather than using a managed endpoint where the platform runs the software. Turn off request and response logging in the server, read the provider’s data-processing terms, and encrypt the disks.
On owned hardware, give the server no outbound internet access once the weights are downloaded.
Either way, load weights only in the safetensors format (it cannot run code when loaded), check the files against the publisher’s checksums, and keep chat history and code inside your own systems.

One caveat for all four

Build the checking into the work. On factual reliability the closed frontier models are still well ahead of every open model, including Kimi K3: Claude Opus 5.5 scores 46 on AA-Omniscience and GPT-6 Astra 44, against Kimi K3’s 20. Since you will not send your work to them, the checks have to come from physics. Test generated solvers against cases with known answers (Couette and Poiseuille flow, Taylor–Green vortex decay, the lid-driven cavity against published reference data). Run grid- and time-step convergence studies, track mass, momentum and energy conservation, and have a second model review the first one’s derivations. A frontier model can still help with generic questions that reveal nothing about your methods.

How this was put together

Benchmarks come from Artificial Analysis (snapshot 1 Oct 2026), at each model’s highest reasoning setting. They are measured at full precision; 4-bit versions may score slightly lower.
CritPt tests research-level physics problems. SciCode tests scientific programming. AA-Omniscience rewards correct answers and penalises confident wrong ones, so it can go below zero. Long context is AA-LCR.
Memory sizes are the actual file sizes of each release on Hugging Face, not parameters × bytes.
Prices: Hugging Face Inference Endpoints price list (1 Oct 2026); B300 cloud and hardware prices from Thunder Compute’s September 2026 price guides; RTX PRO 6000 street price, September 2026. Hardware prices exclude staff time and installation.