Polish language / self-host / sovereignty

Bielik (SpeakLeash)

  • Bielik‑PL‑11B v3 from 04/2026 Apache 2.0

    A tokenizer built for Polish. SpeakLeash and ACK Cyfronet AGH.

  • Minitron 7B / 4.5B v3 7B and 4.5B Apache 2.0

    7B from April 2026, 4.5B from May 2025. When 11B does not fit the card.

  • Bielik Guard from 01/2026 Apache 2.0

    Safety classifier for Polish text, 0.1B and 0.5B.

Bielik 11B v3 from 08/2025 covers 20+ European languages. Cost is GPU, not API tokens.

  • Verified2026-09-20
  • Models covered3
  • Sources5
  • Deployment channels2

Purchase decision (when to choose / when to avoid)

Choose if...

  • You need the best PL in open-weight and want self-host / data sovereignty.
  • You're building RAG or fine-tuning for Polish documents (customer support, government, HR).
  • You want Apache 2.0 license and to avoid vendor lock-in.

Avoid if...

  • You need frontier reasoning/coding at top model level - Bielik is smaller.
  • You want ready enterprise integrations and minimal overhead - SaaS may be faster to start.

Cost in practice (scenarios)

Customer support in Polish

Self-host on smaller GPU cluster; cost depends on SLA and volume.

  • RAG on PL documents
Internal assistant

Very cost-attractive at steady volume, but requires maintenance.

  • SSO, logs, policies
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 (Apache 2.0)
  • Variants and quantizations from HF (vLLM/SGLang)

Data policy

Training on data
Self-host: on your side.
Retention
Self-host: on your side.
Data residency
Full control (on-prem/PL cloud).
Strongest argument is PL + Apache 2.0 + sovereignty.

Enterprise readiness

Admin
On your side (self-host).
SSO/SCIM
On your side.
Audit
On your side.
DPA
On your side (agreements with hosting).
Certifications
Depends on hosting.
Best when you have PL/GDPR requirements and want to avoid vendor lock-in.

Best use cases

  • deployments requiring high Polish language quality (government, customer support, PL documents)
  • sovereign self-host on Polish infrastructure - Apache 2.0, no vendor lock-in
  • RAG and fine-tuning for specific domains in Polish language.

Strengths

  • Best PL quality among open-weight models; custom APT4 tokenizer optimized for Polish.
  • Apache 2.0 - full commercial freedom; developed on Polish HPC (data sovereignty).
  • Bielik Guard - safety classifier for PL (0.1B/0.5B); growing ecosystem.

Weaknesses / risks

  • Smaller scales (4.5B/11B) - weaker reasoning than frontier models; limited multilingual beyond PL/EN.
  • Young ecosystem - fewer ready integrations and documentation than global providers.

Current models (examples)

  • Bielik 11B v3 (08.2025) - multilingual (20+ languages); Bielik-PL-11B-v3.0-Instruct (04.2026) - PL tokenizer.
  • Bielik-PL-Minitron-7B-v3.0 (04.2026); Bielik 4.5B v3 (05.2025) - compact PL.
  • Bielik Guard (01.2026) - safety classifier PL (0.1B/0.5B); Bielik 11B v2 (10.2024).

Alternatives (if this model doesn't fit)

LL
Llama
greater reasoning in open-weight (but community license)
MI
Mistral
EU + model and deployment options
OP
OpenAI
fastest start without MLOps (at the cost of control)