PLLuM

Polish 4B-70B family; base, instruct and chat models

PLLuM (Polish Large Language Model) is a model family developed by a consortium of Polish research and public institutions. The project created a Polish corpus of roughly 140 billion tokens, an instruction dataset and preference data for response alignment. The current second generation (v2.2, model suffix 2512) includes 4B, 8B and 12B models in base, instruct and chat variants, plus 70B instruct and chat models. The first generation also included an 8×7B MoE architecture and non-commercial releases. Licensing varies by repository: current 4B and 12B models use Apache 2.0, while 8B and 70B models use the Llama 3.1 license.

Verified: 2026-06-23

Purchase decision (when to choose / when to avoid)

Choose if...

Polish quality, self-hosting and data control are the priorities.
You are building RAG or an assistant for public-sector, educational or domain documents.
You can evaluate the current 4B, 8B, 12B and 70B variants and select the appropriate license deliberately.

Avoid if...

You cannot review the exact model license; the family includes Apache 2.0, Llama 3.1 and older CC BY-NC 4.0 releases.
You need a turnkey SaaS product with user administration, SLA and per-seat billing.
Frontier coding, multimodality or complex reasoning are the main requirements.

Cost in practice (scenarios)

4B or 12B pilot

GPU infrastructure and operating cost; there is no single official API price for the family.

self-host
RAG or internal assistant
selected quantization and inference engine

70B deployment

Much higher memory and inference requirements; calculate against target traffic and SLA.

multi-GPU infrastructure
monitoring and MLOps team

These are estimates/scenarios (not an invoice). Actual cost depends on context length, number of users, limits and retention policies.

Deployment / data / enterprise

Deployment channels

Self-host from weights published by CYFRAGOVPL on Hugging Face
Deploy with an inference stack compatible with the selected architecture
Choose the current 4B, 8B, 12B or 70B variant according to resources and license

Data policy

Training on data: With self-hosting, data and any further tuning remain under the deployer's control.
Retention: Determined by the organization's self-hosted configuration.
Data residency: Full control with on-premises or selected Polish/European cloud infrastructure.

Model cards and the paper document architectures, corpus and alignment; verify the license per repository.

Enterprise readiness

Admin: Provided by the deploying organization.
SSO/SCIM: Provided by the selected platform or application.
Audit: Provided by infrastructure, application and logging layers.
DPA: Depends on infrastructure-provider agreements.
Certifications: Depends on hosting and deployment design.

The model offers control, but the enterprise operating layer must be built around it.

Best use cases

assistants and RAG systems working mainly with Polish documents
public administration, education and Polish-language research
organizations that want to host a Polish model on controlled infrastructure.

Strengths

A substantial Polish corpus, instruction and preference datasets, and a publicly documented development process.
Multiple model sizes and architectures allow deployment to be matched to infrastructure and workload.
The project includes output correction, filtering, safety evaluation and a public-administration prototype.

Weaknesses / risks

Licensing is not uniform; older variants marked '-nc' are not intended for commercial use.
The 70B variant requires costly infrastructure and MLOps expertise.
Polish specialization does not make the family a direct replacement for frontier models in complex reasoning, coding or multimodality.

Current models (examples)

PLLuM‑4B‑2512 and PLLuM‑12B‑2512 - current base, instruct and chat variants under Apache 2.0.
Llama‑PLLuM‑8B‑2512 and Llama‑PLLuM‑70B‑2512 - current Llama 3.1 variants; 8×7B and '-nc' models are earlier releases.

Alternatives (if this model doesn't fit)

Bielik

another mature Polish open-weight ecosystem with a simpler model lineup

Mistral

European provider with hosted API and open-weight options

Llama

broader global tooling and hosting ecosystem

Summary for decision-makers

Type / role: Polish language / public sector / self-host
When to choose: • assistants and RAG systems working mainly with Polish documents • public administration, education and Polish-language research • organizations that want to host a Polish model on controlled infrastructure.