Updated: November 2026

Polish AI models you can run yourself

Bielik, PLLuM and basal: open-weight models trained on Polish that run on your own server.

  • Open licenses
  • Data stays in your company
  • From laptop to server
Polish flag on an office desk overlooking the Wroclaw market square

At a glance

Four families of Polish models

Each family has a different job. Bielik and PLLuM chat and write, basal makes decisions, and Qra is a base for fine-tuning.

FamilyBuilt bySizesLicenseBest for
BielikSpeakLeash and ACK Cyfronet AGH1.5B, 4.5B, 7B, 11B plus Bielik Guard 0.1B and 0.5BApache 2.0Chat, RAG, Polish writing, content moderation
PLLuMPLLuM consortium, since 2025 HIVE AI, with the Ministry of Digital Affairs4B, 8B, 12B, 8x7B, 70BApache 2.0; Llama-based versions: Llama 3.1Chat, instructions, summaries, public-sector Polish
basalRemigiusz Kinas0.8B to 31B, built on Bielik v3 and Gemma 4Apache 2.0Decisions: choice, yes/no, score with calibration
QraOPI PIB and Gdańsk University of Technology1B, 7B, 13B (base models)Apache 2.0, Llama 2Base for fine-tuning on your own data

All Polish models by RAM

Search for a model and filter by release date, memory or family. We show the newest and the most downloaded release of each model. Last update: November 2026.

Hugging Face snapshot: 2026-11-14

Primary source: Hugging Face model cards

8 GB

18 models

Model Best for Released on HF License Links
basal-2.0-4.5B
License Apache 2.0
Decisions and classification on a small GPU Apache 2.0
basal-2.0-mini
License Apache 2.0
Fastest yes/no and choice decisions Apache 2.0
basal-omni-1.0-mini
License Apache 2.0
Decisions on one 8 GB-class GPU Apache 2.0
basal-omni-2.0-mini-preview
License Apache 2.0
basal 2.0 preview: bidirectional decisions Apache 2.0
basal-omni-pico
License Apache 2.0
Smallest decision model, CPU-friendly Apache 2.0
Bielik-1.5B-v3.0-Instruct
License Apache 2.0
Smallest Bielik for simple tasks and edge devices Apache 2.0
Bielik-4.5B-v3.0-Instruct
License Apache 2.0
Small Bielik for laptops: chat and simple tasks Apache 2.0
Bielik-Guard-0.1B-v1.1
License Apache 2.0
Lightest Polish content moderation Apache 2.0
Bielik-Guard-0.5B-v1.1
License Apache 2.0
Polish harmful-content detection Apache 2.0
Bielik-Minitron-7B-v3.0-Instruct
License Apache 2.0
Lighter 7B Bielik pruned from 11B Apache 2.0
Bielik-PL-Minitron-7B-v3.0-Instruct
License Apache 2.0
Lighter 7B Bielik with a Polish tokenizer Apache 2.0
Llama-PLLuM-8B-chat-2412
License Llama 3.1
PLLuM 8B chat (first release) Llama 3.1
Llama-PLLuM-8B-chat-2512
License Llama 3.1
PLLuM chat built on Llama 3.1 8B Llama 3.1
Llama-PLLuM-8B-instruct-2512
License Llama 3.1
PLLuM built on Llama 3.1 8B Llama 3.1
PLLuM-4B-chat-2512
License Apache 2.0
Light Polish chat Apache 2.0
PLLuM-4B-instruct-2512
License Apache 2.0
Light PLLuM for simple instructions Apache 2.0
Qra-1b
License Apache 2.0
Small base for experiments Apache 2.0
Qra-7b
License Llama 2
7B base for your own fine-tuning Llama 2

basal-2.0-4.5B

Decisions and classification on a small GPU

Released on HF: License: Apache 2.0

basal-2.0-mini

Fastest yes/no and choice decisions

Released on HF: License: Apache 2.0

basal-omni-1.0-mini

Decisions on one 8 GB-class GPU

Released on HF: License: Apache 2.0

basal-omni-2.0-mini-preview

basal 2.0 preview: bidirectional decisions

Released on HF: License: Apache 2.0

basal-omni-pico

Smallest decision model, CPU-friendly

Released on HF: License: Apache 2.0

Bielik-1.5B-v3.0-Instruct

Smallest Bielik for simple tasks and edge devices

Released on HF: License: Apache 2.0

Bielik-4.5B-v3.0-Instruct

Small Bielik for laptops: chat and simple tasks

Released on HF: License: Apache 2.0

Bielik-Guard-0.1B-v1.1

Lightest Polish content moderation

Released on HF: License: Apache 2.0

Bielik-Guard-0.5B-v1.1

Polish harmful-content detection

Released on HF: License: Apache 2.0

Bielik-Minitron-7B-v3.0-Instruct

Lighter 7B Bielik pruned from 11B

Released on HF: License: Apache 2.0

Bielik-PL-Minitron-7B-v3.0-Instruct

Lighter 7B Bielik with a Polish tokenizer

Released on HF: License: Apache 2.0

Llama-PLLuM-8B-chat-2412

PLLuM 8B chat (first release)

Released on HF: License: Llama 3.1

Llama-PLLuM-8B-chat-2512

PLLuM chat built on Llama 3.1 8B

Released on HF: License: Llama 3.1

Llama-PLLuM-8B-instruct-2512

PLLuM built on Llama 3.1 8B

Released on HF: License: Llama 3.1

PLLuM-4B-chat-2512

Light Polish chat

Released on HF: License: Apache 2.0

PLLuM-4B-instruct-2512

Light PLLuM for simple instructions

Released on HF: License: Apache 2.0

Qra-1b

Small base for experiments

Released on HF: License: Apache 2.0

Qra-7b

7B base for your own fine-tuning

Released on HF: License: Llama 2

16 GB

10 models

Model Best for Released on HF License Links
basal-2.0-max
License Apache 2.0
Most accurate basal: routing, rules, scores Apache 2.0
basal-omni-1.0-max
License Apache 2.0
Decisions and scores on Gemma 4 12B Apache 2.0
basal-omni-1.1-max
License Apache 2.0
Decisions over long documents and data Apache 2.0
Bielik-11B-v3.0-Instruct
License Apache 2.0
Most downloaded Polish model: chat and RAG Apache 2.0
Bielik-11B-v3.1-Instruct
License Apache 2.0
Latest Bielik: Polish chat, tool use and reasoning Apache 2.0
Bielik-PL-11B-v3.0-Instruct
License Apache 2.0
Bielik 11B with a Polish-optimised tokenizer Apache 2.0
PLLuM-12B-chat-2412
License Apache 2.0
Polish chat (first release) Apache 2.0
PLLuM-12B-chat-2512
License Apache 2.0
Polish chat assistant Apache 2.0
PLLuM-12B-instruct-2512
License Apache 2.0
Polish instructions, summaries and Q&A Apache 2.0
Qra-13b
License Llama 2
Base for fine-tuning on Polish data Llama 2

basal-2.0-max

Most accurate basal: routing, rules, scores

Released on HF: License: Apache 2.0

basal-omni-1.0-max

Decisions and scores on Gemma 4 12B

Released on HF: License: Apache 2.0

basal-omni-1.1-max

Decisions over long documents and data

Released on HF: License: Apache 2.0

Bielik-11B-v3.0-Instruct

Most downloaded Polish model: chat and RAG

Released on HF: License: Apache 2.0

Bielik-11B-v3.1-Instruct

Latest Bielik: Polish chat, tool use and reasoning

Released on HF: License: Apache 2.0

Bielik-PL-11B-v3.0-Instruct

Bielik 11B with a Polish-optimised tokenizer

Released on HF: License: Apache 2.0

PLLuM-12B-chat-2412

Polish chat (first release)

Released on HF: License: Apache 2.0

PLLuM-12B-chat-2512

Polish chat assistant

Released on HF: License: Apache 2.0

PLLuM-12B-instruct-2512

Polish instructions, summaries and Q&A

Released on HF: License: Apache 2.0

Qra-13b

Base for fine-tuning on Polish data

Released on HF: License: Llama 2

24 GB

1 model

Model Best for Released on HF License Links
basal-omni-1.0-ultra
License Apache 2.0
Strongest basal-omni for hard decisions Apache 2.0

basal-omni-1.0-ultra

Strongest basal-omni for hard decisions

Released on HF: License: Apache 2.0

48 GB

2 models

Model Best for Released on HF License Links
PLLuM-8x7B-chat-2412
License Apache 2.0
PLLuM MoE chat Apache 2.0
PLLuM-8x7B-instruct-2412
License Apache 2.0
PLLuM MoE for instructions and summaries Apache 2.0

PLLuM-8x7B-chat-2412

PLLuM MoE chat

Released on HF: License: Apache 2.0

PLLuM-8x7B-instruct-2412

PLLuM MoE for instructions and summaries

Released on HF: License: Apache 2.0

64 GB

2 models

Model Best for Released on HF License Links
Llama-PLLuM-70B-chat-2512
License Llama 3.1
Largest PLLuM chat model Llama 3.1
Llama-PLLuM-70B-instruct-2512
License Llama 3.1
Largest PLLuM for harder tasks Llama 3.1

Llama-PLLuM-70B-chat-2512

Largest PLLuM chat model

Released on HF: License: Llama 3.1

Llama-PLLuM-70B-instruct-2512

Largest PLLuM for harder tasks

Released on HF: License: Llama 3.1

New

basal: a Polish model that decides instead of writing

basal does not chat. It reads a message, a document or a log and answers a question from a closed list of answers: a choice, yes/no or a score. It returns a probability for each answer, so you know when a decision can be automated and when it should go to a person.

What it is good for

  • Routing tickets to the right team
  • Checking documents against rules and policies
  • Scoring urgency and risk
  • Decisions in production processes, such as downtime calls

Model families

basal-2.0 is built on Bielik v3 (1.5B, 4.5B, 11B). basal-omni is built on Gemma 4 (8B to 31B), and basal-omni-pico (0.8B) on Qwen3.5 and also runs on a CPU. On 14 November 2026 the author released basal-omni-2.0-mini (preview) with bidirectional reading. The full 2.0 family will follow later.

When it makes sense

When to choose a Polish model

Better Polish Correct inflection, style and Polish context in answers and texts.
Data stays in-house The model runs on your server, with no data sent to an API.
Affordable hardware Bielik 4.5B and PLLuM 4B fit in 8 GB of RAM.
Automated decisions basal classifies and scores with a stated confidence.
Content moderation Bielik Guard detects harmful content in Polish.
Open licenses Most models are Apache 2.0, commercial use included.

Comparison

A Polish model or Qwen and Gemma?

Global open-source models handle Polish well too, and the largest ones are stronger at reasoning and coding. Polish models win where natural Polish text, Polish context and a small size matter.

There is no single answer for every company. The best approach is to compare two or three models on your own data, for example Bielik 11B with a model from the same RAM tier in the best local LLMs table.

Hardware

RAM, VRAM, unified memory: what do we mean?

Where the model memory lives depends on your hardware. All tiers in the table assume Q4_K_M quantization.

HardwareWhat to checkExample with Polish models
Mac (M1–M5)Total RAM. CPU and GPU share one memory pool, so no separate GPU is needed.A Mac with 16 GB RAM runs Bielik 11B or PLLuM 12B.
PC with NVIDIA GPUGPU VRAM. The model must fit in the GPU’s memory. System RAM alone is not enough.An 8 GB card (e.g. RTX 4060) is enough for Bielik 4.5B. Bielik 11B needs a 16 GB card (e.g. RTX 5070 Ti).
Multi-GPU serverCombined VRAM. Large models split across GPUs.Llama-PLLuM-70B needs about 48–64 GB: two RTX 5090 cards (2 × 32 GB) or one H100 80 GB.
CPU onlySystem RAM. Slower than a GPU, but small models are usable.basal-omni-pico (0.8B) and Bielik 1.5B run even on an ordinary laptop.

Generation speed depends on hardware and text length, so we do not list it. We explain more in the local LLM ranking.

Step by step

How to run a Polish model locally

  1. 01 Pick a model Match the size to your server or laptop memory using the table above.
  2. 02 Download the weights Get the GGUF version from Hugging Face, ideally Q4_K_M.
  3. 03 Start an engine Ollama or LM Studio for tests, vLLM with an API for production.
  4. 04 Connect it to your business Link the model to your documents, CRM or ticket workflow.

FAQ

Questions about Polish AI models

Is Bielik free and can a company use it?

Yes. Current Bielik models are Apache 2.0, which allows commercial use. The table only lists models whose licenses allow commercial use.

How much memory does Bielik 11B need?

About 16 GB of RAM or VRAM with Q4_K_M quantization. Bielik 4.5B fits in 8 GB.

How is Bielik different from PLLuM?

Bielik is developed by the SpeakLeash community with ACK Cyfronet AGH. PLLuM comes from a consortium of research institutions with the Ministry of Digital Affairs and leans more towards official and public-sector Polish.

Is a Polish model better than ChatGPT?

Not for every task. Large cloud models are stronger at reasoning. A Polish model wins when data must stay in your company and the task needs good Polish.

Is basal a chatbot?

No. basal is a decision model: it picks an answer from the list you give it and returns its probability. It does not write longer texts.

How often do you update the list?

Every month we pull new releases of Bielik, PLLuM, basal and Qra from Hugging Face. Last update: November 2026.

Want to run a Polish model in your company?

Describe the task and the data. We will advise which model to pick, what hardware to run it on and how to connect it to your systems.

  • Reply within 1 business dayA quote, or a few questions if something is missing.
  • A call only if you want one30 minutes, free.
  • ISO 27001:2022EU company, your data stays confidential.
Add details (optional) A few clicks and we can send a quote straight away.
What is it about?
Can the data leave the company?
How many people?
When do you want to start?
Systems to connect

Reply within one business day. No sales call needed.