Which local AI model is best for German?

For German-language tasks, the Gemma models are considered particularly strong, followed by Qwen3 and Mistral. The table shows the German quality per model as an assessment.

What does the context window mean?

The context window is the amount of text, in tokens, that a model can process at once. 32K is enough for most tasks; long documents benefit from 128K or more. More context costs additional memory.

Can these models be used commercially?

That depends on the license (see the “License” column). Apache 2.0 and MIT allow free commercial use; licenses such as the Llama Community License or the Gemma Terms come with restrictions you should verify against the original license when in doubt.

What is an MoE model?

In a Mixture-of-Experts model, all parameters sit in memory, but only a fraction of them (the active parameters) compute per token. This makes it as fast as a much smaller model while requiring the memory of the large one.

Local AI Models: Memory, Context & License at a Glance

Family German quality License

Fits in

38 of 49 models

						License
Kimi K3 New	99	2,800B · 104B active	1,560 GB	1M	good	Kimi K3 License	—	$2.65
The largest openly available model there is: 2.8T parameters, 104B of them active. It needs around 1.5 TB of memory, and there it holds its own against the closed frontier models. Largest open model 1M context Native MXFP4
DeepSeek V4-Pro	98	1,600B · 49B active	880 GB	1M	good	MIT	1457	$1.60
DeepSeek's new flagship: 1.6 trillion parameters, three reasoning tiers and a 1M context for large servers. Frontier class 1M context MIT licence
GLM-5.2 New	97	753B · 40B active	415 GB	1M	good	MIT	—	$0.00
The successor to GLM-5, with five times the context window and an MIT licence, so no usage restrictions at all. Built for long, multi-step tasks. 1M context MIT licence Long tasks
Kimi K2.6	96	1,000B · 32B active	560 GB	256K	good	Modified MIT	1460	$0.95
Moonshot's flagship: one of the strongest open models, for large servers only. Top open weights Frontier class MoE-fast
DeepSeek-V3	95	671B · 37B active	400 GB	128K	good	DeepSeek License	1358	$0.26
A top-tier model for dedicated multi-GPU servers. Frontier class Server hardware
GLM-5	94	744B · 40B active	410 GB	200K	good	MIT	1458	$0.60
Zhipu's frontier MoE for dedicated servers, freely available under the MIT licence. Frontier reasoning MoE-fast MIT licence
DeepSeek V4-Flash	93	284B · 13B active	155 GB	1M	good	MIT	1436	$0.09
The lean V4 variant: frontier reasoning with a 1M context for workstations. 1M context MoE-fast MIT licence
Qwen3 235B-A22B	92	235B · 22B active	140 GB	32K	strong	Apache 2.0	1375	$0.45
A near-frontier MoE model for workstations with plenty of memory. Near-frontier MoE-fast
Kimi K2.7 Code	91	1,000B · 32B active	560 GB	256K	good	Modified MIT	—	$0.71
Moonshot's coding-focused variant of Kimi K2.6: for long-horizon software engineering agents. Coding specialist Agentic Very fast (MoE)
Mistral Medium 3.5 128B	88	128B	75 GB	256K	strong	Modified MIT	—	—
Mistral's dense 128B flagship: instruction-following, reasoning and coding in a single model. Strong all-rounder 256K context EU model
Nemotron 3 Super New	87	120B · 12B active	70 GB	1M	good	NVIDIA Nemotron Open Model	—	—
NVIDIA's open reasoning model, built from Mamba and transformer layers. 12B active parameters keep it fast, and the context window reaches 1M tokens. 1M context Agentic Mamba hybrid
Llama 3.3 70B	86	70B	43 GB	128K	good	Llama Community	1318	$0.10
A large model on par with earlier cloud flagships. Flagship class Highly capable
gpt-oss 120B New	86	117B · 5.1B active	80 GB	128K	good	Apache 2.0	1352	$0.04
OpenAI's large open model, sized to fit exactly one 80 GB card. Apache 2.0, with no restrictions on commercial use. Fits one H100 Tool calling Native MXFP4
Mistral Small 4 119B	85	119B · 6.5B active	70 GB	256K	strong	Apache 2.0	—	$0.08
An MoE model with 119B total and 6.5B active parameters: a general-purpose model and a reasoning model in one. Hybrid reasoning Very fast (MoE) 256K context
Llama 4 Scout	84	109B · 17B active	62 GB	10M	good	Llama 4 Community	1321	$0.10
A massive context window (10M tokens) for entire mountains of documents. 10M context MoE-fast Multimodal
Devstral 2 123B	83	123B	74 GB	256K	good	Modified MIT	—	$0.40
Mistral's large coding agent: 123B parameters for demanding software engineering tasks. Coding specialist Agentic 256K context
Qwen3.6 27B New	82	27B	18 GB	256K	strong	Apache 2.0	—	$0.30
A dense 27B model that beats its own 397B predecessor on coding benchmarks. Currently the best pick for a single 24 GB card. Image & video 262K context Strong at code
Qwen3-Coder-Next	80	80B · 3B active	52 GB	256K	good	Apache 2.0	—	$0.12
An MoE coding model: 80B-class coding quality at 3B speed. Coding specialist Very fast (MoE) Tool calling
Qwen3 32B	78	32B	21 GB	32K	strong	Apache 2.0	1347	$0.08
The most capable model that still fits on a 24 GB GPU. Strongest dense model Reasoning
Gemma 4 31B	77	31B	20 GB	256K	strong	Apache 2.0	1451	—
The Gemma flagship: first-class German, image understanding and a free Apache licence. Top German Multimodal Apache 2.0
QwQ 32B New	76	32.5B	21 GB	128K	strong	Apache 2.0	1336	—
A reasoning specialist that writes out its intermediate steps. Strong at maths and logic, slower to answer because it thinks first. Reasoning Maths & logic Apache 2.0
Gemma 3 27B	74	27B	18 GB	128K	strong	Gemma Terms	1365	$0.08
One of the best options for German-language tasks. Top German Multimodal
Qwen3 30B-A3B	72	30B · 3B active	20 GB	32K	strong	Apache 2.0	1327	$0.12
An MoE model: the quality of a 30B at the speed of a 3B. Very fast (MoE) Reasoning
gpt-oss 20B New	71	21B · 3.6B active	16 GB	128K	good	Apache 2.0	1317	$0.03
OpenAI's open model for consumer hardware. Just 3.6B active parameters keep it fast, and MXFP4 training gets it running from 16 GB. Tool calling Reasoning levels Native MXFP4
Mistral Small 3.1 24B	70	24B	16 GB	128K	strong	Apache 2.0	1303	$0.35
A European model with an excellent speed-to-quality ratio. EU model Fast Tool calling
DiffusionGemma 26B-A4B New	69	25.2B · 3.8B active	14 GB	256K	strong	Apache 2.0	—	—
Google's first open diffusion LLM: generates text in parallel token blocks instead of one token at a time, over 1,000 tokens/sec at 25B-class quality. Diffusion LLM Extremely fast 256K context
Gemma 4 26B-A4B	68	25.2B · 3.8B active	14 GB	256K	strong	Apache 2.0	1438	$0.00
The first MoE model in the Gemma 4 line: 25B-class quality at the speed of a 4B model, with native tool calling. Very fast (MoE) Tool calling 256K context
Devstral Small 2 24B	67	24B	16 GB	256K	good	Apache 2.0	—	—
Mistral's coding agent at 24B: specialised in codebase exploration and multi-file edits. Coding specialist Agentic 256K context
Qwen3 14B	64	14B	11 GB	32K	strong	Apache 2.0	—	$0.12
The sweet spot: strong, versatile, runs on mid-range GPUs. Reasoning Tool calling
Ministral 3 14B	63	14B	11.5 GB	256K	strong	Apache 2.0	—	$0.20
The sweet spot of the Ministral 3 line: strong German, image understanding and 256K context. Multimodal (image) 256K context Strong German
Gemma 4 12B	62	12B	9.5 GB	256K	strong	Apache 2.0	—	—
The new Gemma generation: strong German and image understanding, now under Apache 2.0. Strong German Multimodal 256K context
Phi-4 14B	60	14B	11 GB	16K	solid	MIT	1256	$0.07
Microsoft's specialist for logic, maths and programming. Reasoning Maths & code
Gemma 3 12B	58	12B	9.5 GB	128K	strong	Gemma Terms	1342	$0.05
Outstanding German and image understanding at moderate requirements. Strong German Multimodal
GLM-4.1V 9B Thinking New	56	10B	8.5 GB	64K	solid	MIT	—	—
A vision-language model with a built-in reasoning mode: matches 72B-class models on image and document understanding. Vision (image) Reasoning MIT licence
Qwen3-VL 8B	54	8B	7.5 GB	256K	good	Apache 2.0	—	$0.12
Sees images and documents: Qwen3 with vision for local OCR and analysis. Vision (image) Long context Tool calling
Ministral 3 8B	53	9B	7.5 GB	256K	strong	Apache 2.0	—	$0.08
A versatile 8B model from Mistral's new Ministral 3 line with image understanding and a huge context window. Multimodal (image) 256K context Tool calling
Qwen3 8B	52	8B	7 GB	32K	strong	Apache 2.0	—	$0.12
An excellent all-round model for everyday enterprise use. Reasoning Tool calling Multilingual
EuroLLM 22B New	48	22.6B	15.5 GB	32K	strong	Apache 2.0	—	—
A larger EuroLLM generation from the UTTER project with 22.6B parameters. Unlike the 9B version and Teuken, it ships with a 32K context window. 35 languages Made in Europe 32K context
Llama 3.1 8B	46	8B	6.5 GB	128K	solid	Llama Community	—	—
The classic, with a vast ecosystem of specialised variants. Widely adopted Many fine-tunes
Gemma 4 E4B	41	8B	4.5 GB	128K	good	Apache 2.0	—	—
A compact multimodal Gemma 4 model: image and audio input at a low memory footprint thanks to Per-Layer Embeddings. Efficient (PLE) Image & audio 128K context
Qwen3 4B	38	4B	4 GB	32K	good	Apache 2.0	—	—
A strong entry point that already runs smoothly on small devices. Great value Tool calling
Ministral 3 3B	37	3.8B	4.5 GB	256K	good	Apache 2.0	—	$0.10
A compact Mistral model with image understanding and a large context window for low-end hardware. Multimodal (image) 256K context Small
Gemma 3 4B	36	4B	4.5 GB	128K	good	Gemma Terms	1303	$0.05
Understands images and offers a large context window. Multimodal (image) Long context
EuroLLM 9B New	34	9.2B	7.5 GB	4K	strong	Apache 2.0	—	—
A European multilingual model from the UTTER project, trained on every official EU language. Held back by the same 4K context as Teuken. 35 languages Made in Europe Apache 2.0
Gemma 4 E2B	33	5.1B	3 GB	128K	good	Apache 2.0	—	—
A tiny edge model from the new Gemma 4 generation with image and audio understanding thanks to Per-Layer Embeddings. Efficient (PLE) Image & audio For low-end hardware
Llama 3.2 3B	30	3B	3.5 GB	128K	solid	Llama Community	1167	$0.05
A lightweight model with a large context window. Long context Edge-ready
Teuken 7B New	30	7B	6 GB	4K	strong	Apache 2.0	—	—
Trained by Fraunhofer IAIS in the OpenGPT-X project, centred on German and the other EU languages. A 4K context window and a 2023 knowledge cut-off limit what it can do today. All 24 EU languages Made in Germany Apache 2.0
Qwen3 1.7B	24	1.7B	2.5 GB	32K	good	Apache 2.0	—	—
A compact all-rounder with decent German. Small Multilingual
Gemma 3 1B	18	1B	2 GB	32K	good	Gemma Terms	—	—
A tiny model for simple tasks and low-end hardware. Very small Blazing fast

LMArena (as of 09/13)

Last extended on 2026-08-10. Every value is checked against its official model card; the oldest such check dates from 2026-06-09.

What changed recently

2026-08-10: EuroLLM 22B
2026-08-05: Teuken 7B, EuroLLM 9B, GLM-4.1V 9B Thinking, DiffusionGemma 26B-A4B, gpt-oss 20B, Qwen3.6 27B, QwQ 32B, Nemotron 3 Super, gpt-oss 120B, Kimi K3, GLM-5.2

How the numbers come together

Transparency first: these figures are carefully estimated reference values for real-world use, not lab benchmarks. The verification date sits below the table; the icon next to each model links to its official model card.

Memory requirement

The memory value applies to the Q4_K_M quantization (the default in Ollama and LM Studio) including a small context window. Rule of thumb: around 0.55 GB per billion parameters plus some overhead. Models trained quantisation-aware from the start (gpt-oss, Kimi K3) carry the vendor’s own published minimum instead, or the size of the published files where the vendor never published one.

Mixture-of-Experts

For MoE models, the total parameter count determines the memory requirement (all experts sit in RAM), while only the active parameters determine the speed. Hence the speed advantage at the same size.

Strength score

The strength is a rounded reference value (0–100) that increases with model size. It helps when comparing within this list, but it is not an official benchmark score.

Note on the context window: the Qwen3 text models from 4B upwards run natively with 32K tokens and can be extended to up to 128K via YaRN (the 1.7B model stays at 32K). The table shows the native value.

Frequently asked questions

Which local AI model is best for German?: For German-language tasks, the Gemma models are considered particularly strong, followed by Qwen3 and Mistral. The table shows the German quality per model as an assessment.
What does the context window mean?: The context window is the amount of text, in tokens, that a model can process at once. 32K is enough for most tasks; long documents benefit from 128K or more. More context costs additional memory.
Can these models be used commercially?: That depends on the license (see the “License” column). Apache 2.0 and MIT allow free commercial use; licenses such as the Llama Community License or the Gemma Terms come with restrictions you should verify against the original license when in doubt.
What is an MoE model?: In a Mixture-of-Experts model, all parameters sit in memory, but only a fraction of them (the active parameters) compute per token. This makes it as fast as a much smaller model while requiring the memory of the large one.

From model to production system

Run your model locally and Corporate LLM turns it into a production system: RAG, agent system, skills, and connectors. 100% GDPR-compliant.

Start for free Back to the calculator