Kimi K3, Soofi S and Apertus: an open-model snapshot

The summer of 2026 has brought another wave of open-weight models. Moonshot AI published Kimi K3, while Germany’s publicly backed Soofi project presented first results without releasing a production assistant for general use yet. The proprietary providers updated their models at the same time. Rankings give you a useful snapshot and a fragile one. Analyst Nathan Lambert puts the open-weight lag at “3 to 5 months rather than 6 to 9”.

Outside the laboratories, the decision has four parts: the quality on your tasks, the release and licence status of the model, the full path your data takes, and the cost to operate it.

A July 2026 snapshot

The chart records an Artificial Analysis snapshot taken in July 2026. The versions, the scores, the prices and the methods change; it isn’t a permanent result, and it isn’t a test of your workload either.

You can read that snapshot three ways. The gap at the top of the aggregate index is small, while larger gaps remain on some coding tasks. The price gaps can be substantial too, but a fair comparison needs the same region, the same lengths, the same treatment of the cache, the same output and the same service level. And Apertus and Soofi pursue openness and independence goals that don’t fit into a single benchmark row.

Kimi K3: the open model in the top league

Kimi K3’s official repository describes 2.8 trillion total parameters and a context of one million tokens. External evaluations place it near the proprietary models on some tasks; that doesn’t predict what it will do on your documents, your languages or your tools.

“Open weights” doesn’t mean “on your premises”: an estimate of roughly 1.4 TB assumes weights at around four-bit precision, and it should say so. The runtime, the cache, the replication and the storage come on top. A pragmatic route may be elsewhere: the predecessor Kimi K2.6 is served by Infomaniak in Geneva, with the processing in Switzerland and a published commitment not to train on your requests. Those assurances have to end up in the contract, next to the access, the subprocessors, the logs and the exit terms.

Soofi S: the German surprise

Soofi S comes from a German consortium with Fraunhofer, DFKI and universities in it. Fraunhofer announced first results in June 2026, but the model card describes the current checkpoint as a closed-beta research artefact. It’s a base model. Nobody has instruction-tuned it or aligned it to serve as a production assistant.

In practice: Soofi S is a project to watch and to evaluate in a research setting, not an internal assistant to deploy as it stands. A future instruction-tuned checkpoint should be tested in German and in French, for safety and under the real workload.

Apertus: the transparent Swiss model

Apertus (EPFL, ETH Zurich and CSCS) belongs to a small group that pursues broad openness: the weights, the code, the method and information about the training data. It isn’t the world’s only large open model, and it isn’t necessarily the strongest on raw benchmarks.

In practice: Apertus deserves evaluation when documentation and Swiss context come first: administrations, towns, schools, any project where every component has to be accounted for. Whether it suits fine-tuning depends on the task and the dataset. The 8B variant can run within 16 GB with the right quantisation and runtime; the context and the operating system also take memory.

Mistral: the European, almost fully open

Mistral publishes models under different licences and release terms. Devstral Small 2 is a candidate to review confidential code. Whether it fits on one 24 GB card depends on the quantisation, the context, the runtime and the tools; measure it rather than calling the choice obvious.

Running them

OpenRouter aggregates many models behind one key and makes a first comparison convenient. But your requests pass through this New York company on their way to the selected providers. Testing with synthetic data may be fine; a patient, pupil or resident file first needs contractual, legal and technical qualification of the route you select. A Swiss-hosted alternative exists: Infomaniak AI Services serves Apertus, Kimi K2.6 and Mistral Small 4 from Switzerland, with an OpenAI-compatible API.

Locally, Ollama installs a model in one command, with one trap to know: the models marked “cloud” in its library run on Ollama’s servers, not on yours. To serve a whole team, the standard is vLLM. Three complete configurations are set out in the local-AI machines article.

The return on running locally

The electricity for a million tokens may cost less than one franc in a fast, well-used configuration, but that isn’t a constant. Publish the power draw, the throughput, the electricity price, the utilisation and the elapsed time. The total cost includes the hardware, the operations, the cooling, the security, the failures and the replacement. At low volume an API is often cheaper; running locally may win at scale, or when control is the requirement, after you do the sums.

The other gains depend on the complete stack. Running locally can remove certain subprocessors and let you work offline, but only when the runtime, the telemetry, the updates, the support, the identity and the backups are also under control. A downloaded model may stay available when the APIs retire theirs. And reproducing the same result in five years means pinning the model file, the quantisation, the runtime, the dependencies, the drivers, the prompt templates and the sampling parameters. Permanence is an archival practice, not a property of the weights alone.

How to decide

I never choose a model from a leaderboard. I test it on the client’s own corpus, with their actual questions and their actual permissions.

Build a representative evaluation set and version it. For every model, write down the exact identifier, the licence, the precision, the access method, the language quality, the error rate, the time to first token, the throughput, the context, the input and output price, the cache, the operator, the locations, the subprocessors and the date you measured. Then answer three questions: can we run it, is it good enough on our test set, and can we operate it safely and lawfully?

The snapshot behind the chart

• Moonshot AI: official Kimi K3 repository
• Fraunhofer: initial Soofi results
• Soofi S Base model card
• Apertus 70B model card

I took the release status, the catalogue and the comparisons on 14 August 2026. Scores and prices move from one month to the next, and so do the models themselves. The chart is an external snapshot; the decision should rest on an internal evaluation you can reproduce.