Best open-source LLMs in 2026 for business use: Qwen3.8, GLM 5.3, Gemma 4 and Nemotron compared

What makes an open-source LLM good for business?

Leaderboards rank models on a single number. A business has five questions, and the benchmark is only one of them: can I ship this under its licence, does it speak my customers’ languages, does it fit my documents, can it call my tools, and what does a million tokens cost?

Open weights change the last question. You can run the model on your own GPUs or pay an API per token, and because the model is the same, you can move between the two without rewriting your application. That is why most of this article is about the models themselves; only the last section is about one way of running them.

We picked the four families that cover the common jobs: Qwen3.8 from Alibaba, GLM 5.3 from Z.ai, Gemma 4 from Google DeepMind and Nemotron 3.5 Lightning from NVIDIA. ALIA 40B, the Spanish public model, appears where language matters.

The four families at a glance

Sizes, licences and languages as each official model card states them. Context is the window the card gives for the original model.

Two of these are mixture-of-experts models that only activate a fraction of their weights for each token: Gemma 4 26B A4B uses 3.8B of its 25.2B parameters, and Nemotron 3.5 Lightning 3B of its 30B. That is what makes them cheap to serve for their size, and it is why their price per token sits close to that of much smaller models.

What the benchmarks say, from the official model cards

Every figure below is copied from the maker’s own model card on Hugging Face, checked on 9 November 2026. Each vendor chooses its own benchmarks and runs them with its own settings and harness, so a dash means the card does not publish that test, not that the model failed it. Read across rows rather than down columns: GPQA Diamond and Terminal Bench 2.1 are the two tests where three of the four publish a score.

Benchmarks copied from the official model cards on 9 November 2026. Sources: huggingface.co/zai-org/GLM-5.3, huggingface.co/Qwen/Qwen3.8-27B, huggingface.co/google/gemma-4-26B-A4B-it, huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4

What the table supports: GLM 5.3 is ahead on long agentic work in a terminal and on real repositories (Terminal Bench 2.1 and DeepSWE), which is what Z.ai built it for. Qwen3.8 27B posts the highest GPQA Diamond and LiveCodeBench v6 scores of the group with a fraction of GLM’s parameters. Gemma 4 and Nemotron 3.5 Lightning give up some accuracy in exchange for very few active parameters.

Two caveats. First, these scores are for the original weights, except Nemotron’s, which NVIDIA publishes for the 4-bit NVFP4 checkpoint itself; its card shows both versions side by side, 81.94 on MMLU Pro in BF16 and 81.62 in NVFP4. Second, Qwen3.8 is a vision-language model on its card, and its vision scores only apply where it is served with images enabled.

Which model for which job

If you only read one table, read this one. It maps the common business jobs to the model that fits them, using what the cards say and the capabilities each model has.

A rule that saves money: start with Qwen3.8 27B, drop to Qwen3.5 9B when the task is simple and the volume high, and move up to GLM 5.3 only when the smaller model fails on your own examples. Test on twenty real prompts from your product before you trust any leaderboard, including the one above.

Licences: what you can actually ship

Apache 2.0 (Qwen3.8, Qwen3.5, Gemma 4 and ALIA): you can use the model commercially, modify it and fine-tune it, as long as you keep the licence and notices. It is the least friction a legal team will see.

GLM 5.3 ships under its own licence. The permissions read like MIT, with one added condition: a model-as-a-service business whose group revenue exceeds 10 billion US dollars in any twelve months must pass a security review by Z.ai before commercial use. Building the model into your own end-user product does not count as model-as-a-service under that licence.

Nemotron 3.5 Lightning uses the OpenMDW License Agreement, version 1.1, and NVIDIA’s card states that the model is ready for commercial use.

None of this is legal advice. Read the licence file that sits next to the weights before you ship, and check the version: licences can change between releases of the same family.

Licences as stated on the official model cards on 9 November 2026. Sources: huggingface.co/zai-org/GLM-5.3, huggingface.co/Qwen/Qwen3.8-27B, huggingface.co/google/gemma-4-26B-A4B-it, huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4, huggingface.co/BSC-LT/ALIA-40b-instruct-2606

Why is Llama not on this list?

We do not serve Llama, and European companies have a licensing reason to look twice at it: Meta’s Acceptable Use Policy for Llama 4 does not grant the licence rights for its multimodal models to individuals domiciled in the EU or companies with their principal place of business there, and the Llama 4 models are natively multimodal according to their card. We compare Qwen, Llama and Gemma on European languages, with the sources, in a separate article.

How this works on GPU Flow

Token Factory serves these models through an OpenAI-compatible API at https://api.gpuflow.ai/v1: you change the base URL and the model name, and keep your code. Billing is per million tokens, in euros, from a prepaid balance. These are today’s prices, read from the same source as our pricing page, with VAT included:

Model pages: GLM 5.3 NVFP4, Qwen3.8 27B, Qwen3 14B, Qwen3.5 9B, ALIA 40B Instruct, Gemma 4 26B A4B, Nemotron Nano 12B VL, Qwen3-VL 32B OCR, Nemotron 3.5 Lightning 30B

The context we serve can be shorter than the native one, because smaller models run on a MIG slice of an NVIDIA B200. Qwen3.8 27B runs with 98,304 tokens and as text only. Gemma 4 runs with 32,768. GLM 5.3 accepts up to 1,048,576 tokens, but a key is limited to 256,000-token prompts by default; ask us if you need more.

GLM 5.3 NVFP4 is served outside our cluster in Spain; the other models run on our own B200 GPUs. Nothing you send is used to train any model.

New accounts get 5 € of free Token Factory credit to compare models on their own prompts. A card is required to sign up, and when the balance runs out the API returns an error instead of charging the card.

Related

• Qwen vs Llama vs Gemma: which open model for Spanish and European languagesWhich open model handles Spanish and European languages: what the Qwen, Llama and Gemma model cards support, the Llama 4 EU clause, and ALIA for Catalan.
• Migrate from the OpenAI API to an open-source alternative in 15 minutes (code included)Move from the OpenAI SDK to GPU Flow in five steps: change the base URL, pick the equivalent model and keep streaming, tools and reasoning.