Open-weight model for German B2B: Qwen, Kimi, Nemotron, or Minimax?

The license is the first gate, because it decides every commercial on-prem use before a single benchmark counts.

Three of the four are straightforward here. Qwen3 is under Apache 2.0, Minimax M2 under MIT, Kimi K2 under a modified MIT license (Hugging Face). These are real, free grants: commercial use, redistribution, fine-tuning, no phoning home.

🔸 Nemotron is the special case.
NVIDIA's Open Model License is not a pure Apache or MIT license. It ties usage to NVIDIA's conditions and terminates automatically if you remove the built-in safety mechanisms (NVIDIA). The model is good, but this clause lands on the legal department's desk first, not engineering's. We described NVIDIA's approach in detail in our piece on Nemotron 3.

Anyone using sovereignty as an argument should be able to read the license like a contract, because that is exactly what it is.

The honest answer: there is no clean four-way comparison of these models in German, so you have to build it yourself.

German eval infrastructure exists. SuperGLEBer, the Occiglot Euro-LLM leaderboard, and MMLU-ProX across 29 languages cover German (ACL Anthology). What is missing is a published head-to-head comparison of exactly these four on German text. Qwen3 was trained on 119 languages and is the family with the broadest documented multilingual coverage (Qwen); for Kimi, Nemotron, and Minimax, the providers do not name German as a particular focus.

The practical consequence is uncomfortable but clear: benchmark rankings on English tests say little about quality on your German contracts, proposals, or support texts. The only reliable test is the one on your own material. How to read leaderboards at all without being misled is covered in the piece on the local model.

Hardware fit sorts the field faster than any benchmark, because a model that does not fit on your cards simply falls out of the running.

Qwen3 is the exception here. The dense variants from 4B to 32B run quantised on one to two GPUs, and the 30B-A3B MoE variant delivers model quality with only around 3B of active compute. That is the profile that fits a typical on-prem box in the Mittelstand.

Kimi K2 and Minimax M2 are frontier MoE models. Minimax activates only 10B parameters, but the full weights require a multi-GPU server; Kimi, with 1 trillion total parameters, even more so. Nemotron 3 Nano is very lean at 3.6B active parameters and runs on a single card, but falls back due to its license. What this VRAM reality actually costs, we worked out in the hardware piece.

With open-weight models, the sovereignty question shifts: the weights run entirely offline, so there is no remote kill switch. What remains is whatever is baked into the model itself.

Three of the four labs are Chinese (Qwen, Kimi, Minimax), one is American (Nemotron). Because the weights run locally, the real question is not the server location but the alignment built into the model itself. Independent tests report high refusal rates for Chinese models on politically sensitive topics, strongest in Chinese and weaker but still present in German (ChinaBench). These figures are self-published and not peer-reviewed, so they should be read with caution, but the effect is real enough to warrant an honest paragraph in your requirements specification.

For pure B2B text work, meaning drafts, extraction, classification, and retrieval over your own documents, this effect is usually manageable. It is a question of output quality and external perception, not a security hole. We draw the larger sovereignty arc, European models, and what the label is actually worth, in our piece on sovereign European AI.

The decision is not patriotic and not driven by the leaderboard, it follows a sequence. Here is how I do it:

1. Check the license. Only models with a genuine Apache or MIT grant make the shortlist without a legal query. That means Qwen3, Kimi K2, and Minimax M2 here.
2. Test on your own German. Not the English benchmark, but run fifty of your real texts against two candidates and evaluate them blind.
3. Match the hardware. Whatever fits on your one to two cards almost always wins, because operation otherwise becomes expensive and fragile.

For most Mittelstand companies, that lands on Qwen3 as the starting point: freely licensed, broadly multilingual, in the right hardware window. Kimi and Minimax are the choice when the server is already in place and reasoning or agents are the priority. We took a dedicated look at the open-weight coding model Kimi K2.7.

And, as always with us: build the architecture so you can swap the model out. The next better open-weight model is certainly coming. The real question is not which one leads today, but which of your tasks actually need the strongest model, and which only need a good one that belongs to you. Which of your texts would you entrust to a model you have not tested yourself?