{"id":21285,"date":"2026-10-09T17:23:00","date_gmt":"2026-10-09T15:23:00","guid":{"rendered":"https:\/\/www.lukaswojcik.com\/blog\/?p=21285"},"modified":"2026-10-05T16:49:57","modified_gmt":"2026-10-05T14:49:57","slug":"freie-llm-modelle-de-en-pl-en","status":"publish","type":"post","link":"https:\/\/www.lukaswojcik.com\/blog\/en\/cloud-ai\/freie-llm-modelle-de-en-pl-en\/","title":{"rendered":"Freely Available LLMs Compared: Qwen, Gemma, Ministral, Llama and DeepSeek for DE, EN and PL"},"content":{"rendered":"

A freely available language model becomes useful locally only when task, language and hardware align. A strong general benchmark does not establish Polish writing quality or the memory requirements of a specific quantisation. Several comparison axes are therefore more useful than a single leaderboard.<\/p>\n

This selection covers documented families and concrete candidates. It does not report an original quality benchmark. Download access, licensing and execution requirements receive separate treatment.<\/p>\n

\"Five
Separate documented properties from original quality evaluation.<\/figcaption><\/figure>\n

Candidate classes<\/h2>\n\n\n\n\n\n\n\n\n\n
Candidate<\/th>\nUseful comparison<\/th>\nSpecific check<\/th>\n<\/tr>\n<\/thead>\n
Qwen3.5-9B<\/td>\nCompact general assistant<\/td>\nHybrid architecture and backend<\/td>\n<\/tr>\n
Gemma 4 Instruct<\/td>\nMatching size class and multimodality<\/td>\nExact variant rather than family<\/td>\n<\/tr>\n
Ministral 3 8B\/14B<\/td>\nLocal chat and structured output<\/td>\nArtifact precision and language test<\/td>\n<\/tr>\n
Llama 3.3 70B Instruct<\/td>\nLarger reference model<\/td>\nWeights, context and licence<\/td>\n<\/tr>\n
DeepSeek-R1-Distill<\/td>\nReasoning in a suitable size class<\/td>\nBase model and output workload<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

Availability is an incomplete description<\/h2>\n

Downloadable weights can still carry particular usage conditions. The specific model card and included licence determine the relevant terms. Derivatives can inherit conditions from their base model. A repository name does not settle that assessment.<\/p>\n

Qwen3.5-9B and the selected Ministral 3 14B Instruct entry identify Apache 2.0. This does not assign the same licence to unrelated families or older generations. Base, instruction-tuned and reasoning variants also differ despite similar names.<\/p>\n

DE, EN and PL need separate results<\/h2>\n

A useful task set includes summarisation, editing, JSON extraction, code analysis and answers grounded in a supplied document. Equivalent content is expressed naturally in all three languages. Assessment counts factual errors, omissions, format violations and unsupported additions.<\/p>\n

A translated English benchmark does not replace a Polish language evaluation. A vendor list of supported languages does not demonstrate equal quality either. Names, inflection and neutral editorial style can produce distinct Polish failure patterns. Separate result columns preserve those differences.<\/p>\n

Separate model-class and memory-budget comparisons<\/h2>\n

One comparison group keeps model class similar. Another keeps available memory fixed. This reveals whether a smaller higher-precision model or a larger more aggressively quantised model better serves the actual task.<\/p>\n

Active parameters account for only part of an MoE model’s requirements. Total weights remain relevant to capacity. Llama 4 Scout and large current DeepSeek models do not become small desktop models merely because active-parameter counts sound smaller. Offloading further changes response behaviour.<\/p>\n

A reproducible shortlist<\/h2>\n

Each result needs model ID, revision, quantisation, runtime, context, chat template and sampling settings. Reasoning output also requires a recorded generation limit. Otherwise a time comparison may simply measure longer answers. Image tasks remain a separate evaluation so that text comparisons do not mix different capabilities.<\/p>\n

The existing memory estimator helps with initial selection. A small local quality test with verifiable expected answers follows. Only then does a defensible shortlist emerge for German, English or Polish. A universal ranking of all freely available models would be too coarse for that decision.<\/p>\n

Memory planning for model selection: LLM VRAM estimator.<\/a><\/p>\n

\n

Related tools<\/h2>\n