Bielik: the Polish Open-Source LLM for Business. Model Family, Licence, Hardware and Honest Benchmarks (2026)

Bielik is an open family of Polish language models built by the SpeakLeash foundation and ACK Cyfronet AGH, released entirely under Apache 2.0, so it is free to use commercially with no licence fees. The variants run from 1.5 to 11 billion parameters, and the 11B version quantized to 4-bit fits on a single 24 GB card. A Syntalith implementation starts from €3,500 net.

Quick answer

Bielik is the simplest open Polish model to stand up on your own hardware: the licence attaches no business conditions and the authors publish quantized builds for the popular runtimes themselves. Five facts carry the decision:

• Licence: Apache 2.0 across the whole family we checked (HuggingFace model cards, 7 August 2026). Commercial use, modification, deployment at a client, no revenue threshold and no vendor acceptable-use policy.
• Hardware: a 24 GB card is enough for 11B at 4-bit. That is arithmetic you can redo yourself: roughly 0.5-0.6 GB per billion parameters plus 15-25% headroom for KV cache and context.
• Quality: it depends on the task. SpeakLeash's own leaderboard puts the 11B v2.3 at a 65.71 average. The independent Oxido test reported by Bankier.pl on 17 March 2026 placed Bielik in the lower half of the field. Both numbers are true and they measure different things.
• Context: 32,768 tokens is confirmed only for the v2 series. The v3 cards do not state a figure, so we do not assume one.
• Syntalith implementation cost, net: AI automations from €3,500, AI apps and AI agents from €6,000, typical full builds €6,000–35,000, maintenance priced individually.

The start is free: a process scan (€0) is a 30-minute engineer call plus a written takeaway within two business days. No sales deck; you talk to the engineer who would build it. The wider view of both Polish model families is in Polish language models: PLLuM and Bielik for business.

Who builds Bielik and where was it trained?

Bielik is a joint effort of two organisations: the SpeakLeash foundation (also known as Spichlerz), which runs the project in the open, and ACK Cyfronet AGH, the academic supercomputing centre in Kraków. That attribution appears on every model card on HuggingFace (read 7 August 2026).

Training runs on the Athena and Helios supercomputers at Cyfronet under PLGrid computational grants; the Bielik-11B-v3.0-Instruct card names grant PLG/2024/016951. The best-documented run is Bielik-4.5B-v3: trained on Helios across 256 NVIDIA GH200 accelerators, on 292 billion tokens for 1.2 epochs, initialized from Qwen2.5 3B, using SpeakLeash's own framework, ALLaMo.

This matters to a buyer for two reasons. The compute budget is public, so the scale is visible: an academic and community project rather than a frontier lab. And the infrastructure is Polish and public, which makes provenance easy to document in organisations that have to document it.

A commercial layer has grown around the project. The president of Bielik.AI is Sebastian Kondracki, who told Bankier.pl (28 June 2026) that the models had reached 2.8 million downloads and named healthcare and finance as sectors where data cannot be sent to cloud servers. That figure comes from a single press source and we treat it accordingly.

What is in the Bielik family in 2026?

The family is broader than the name suggests. The state below is as of 7 August 2026, from the model cards and the SpeakLeash blog.

The v3 series shipped in 1.5B, 4.5B and 11B sizes. Two 2026 dates are worth remembering:

• 15 April 2026: Bielik-PL-11B-v3.0-Instruct and Bielik-PL-Minitron-7B-v3.0-Instruct, the first Bielik variants with a dedicated Polish-optimized tokenizer. They come with the paper "Advancing Polish Language Modeling through Tokenizer Optimization in the Bielik v3 7B and 11B Series" (arXiv 2604.10799, 12 April 2026). The authors report that the PL-tokenizer models preserve performance across nine Polish and multilingual benchmarks and surpass the originals on CPTUB and Polish EQ-Bench, with English capability largely intact.
• 24 June 2026: Bielik-11B-v3.0-DFlash, a draft model of around 1B parameters for speculative decoding. In practice the small model proposes a run of tokens and the large one verifies it, which cuts latency without changing the main model.

A v3.1 release extending coverage from 30 to 50 languages was announced for June 2026 (Bankier.pl, 28 June 2026), but as of 7 August 2026 we did not see it on the SpeakLeash model listing. We treat it as announced and unconfirmed.

Those two make more practical difference than another benchmark point: the tokenizer lowers the token count for the same Polish sentence, DFlash lowers latency. Both feed straight into the cost of one query on your GPU.

What does the Apache 2.0 licence actually buy you?

This is Bielik's strongest argument. Apache 2.0 lets you use the model commercially, modify it, fine-tune it on your own data and deploy it at a client, with no licence fees, no user-count threshold and no vendor-imposed acceptable-use policy. Every Bielik variant we checked carries that licence. The contrast with PLLuM shows why this is more than a formality:

Source: HuggingFace model cards and organisation listing plus pllum.org.pl, read 7 August 2026.

The PLLuM catch runs deeper, because the -nc- variants saw roughly 150 billion tokens of Polish text in continued pretraining while the openly licensed variants saw only 28-30 billion (a copyright-clean subset). The best-trained versions are therefore the ones you may not ship in a product. Bielik has no such tradeoff: the version you can deploy at a client is the same one the authors benchmark. How PLLuM looks from the deployment side is covered in deploying PLLuM in a company.

One caveat: a model licence is not GDPR compliance. Open weights settle the right to use, leaving the processing agreement, data location and records of processing untouched. That is separate work, covered in GDPR and the DPA when deploying AI.

How much hardware does Bielik need?

What follows is arithmetic you can redo yourself, rather than a measured result. The rule of thumb: in bf16 a model takes roughly 2 GB per billion parameters, at 8-bit around 1 GB, at 4-bit around 0.5-0.6 GB. Add 15-25% for KV cache, activations and CUDA context.

The conclusion is simple: 11B at 4-bit is around 7 GB of weights and runs comfortably on a single 24 GB card. For comparison, PLLuM's own infrastructure guidance (pllum.org.pl, read 7 August 2026) puts a 12B generator at around 48 GB, because it assumes fp16 serving with production headroom plus a separate retrieval and reranker stack at 24-48 GB. Both figures are honest; they describe different deployments.

Quantization has a price, and precision matters here: no published, dated measurement of Bielik's quality loss at Q4 versus bf16 exists, so nobody should be quoting you a percentage. SpeakLeash's claim of around 90% quality retention applies to Minitron, meaning pruning and distillation rather than quantization. If a vendor blends the two, you now know what the rest of their numbers are worth.

Getting it running is a short path. SpeakLeash publishes official quantizations: for Bielik-11B-v2.3-Instruct those are GGUF (Q4_K_M, Q5_K_M, Q6_K, Q8_0), GPTQ 4-bit and FP8 for Ada Lovelace and Hopper cards, plus experimental IQ variants with imatrix calibration. Bielik-11B-v3.0-Instruct ships 17 quantized variants targeting llama.cpp, LM Studio, Jan and Ollama, and the official Ollama namespace is ollama.com/SpeakLeash. The architectures are standard Mistral and Qwen derivatives, so vLLM serves them the way it serves their base models, though SpeakLeash does not name vLLM as a supported target.

The hosting arithmetic, as a modelled scenario: a 24 GB-class card in an EU data centre is the cheapest sensible tier for Bielik 11B at 4-bit. If you need headroom for longer context, parallel sessions or a 70B model from another family, Hetzner's GEX131 in Germany, with an RTX PRO 6000 Blackwell Max-Q and 96 GB of VRAM, starts at €889 per month excluding VAT (Hetzner press release, 11 December 2025), with an optional GDPR Article 28 data processing agreement and no minimum term. Substitute your own numbers: monthly server cost divided by query count gives a unit cost you can hold against an API provider's token rate. The full hardware, power and maintenance calculation is in deploying a Polish LLM: hardware, costs, GDPR.

How well does Bielik actually score?

Care is needed here, because two sets of numbers circulate and both are true.

The authors' numbers. Bielik-11B-v2.3-Instruct posts a 65.71 average on the Open PL LLM Leaderboard, and Bielik-11B-v3.0-Instruct posts 65.93 at 5-shot, with Bielik-PL-Minitron-7B-v3.0-Instruct at 61.66 (model card and the speakleash.org blog of 15 April 2026). SpeakLeash also states that v3.0-Instruct beats Meta-Llama-3.1-70B-Instruct on that battery. The leaderboard is maintained by SpeakLeash itself, and that has to be said plainly. It does not make the numbers false; it means they come from an interested party and measure a battery of 29 Polish NLP tasks rather than open-ended reasoning, code or adherence to complex instructions.

The independent numbers. A test by the firm Oxido, reported by Bankier.pl on 17 March 2026, covered 20 tasks across 10 categories: writing emails, business advice, language correctness, Polish history and culture (including reciting the first twelve verses of the invocation to Pan Tadeusz), legal and tax questions, and marketing. A Google model won, with Qwen and Meta's Llama also on the podium, and Bielik and PLLuM finished in the lower part of the ranking. On the Pan Tadeusz task Bielik came eighth and PLLuM third from last. The test's author, Marek Jeleśniański, called Bielik's placement a decent result given incomparably smaller resources.

The two pictures reconcile. Polish benchmarks measure classification, reading comprehension and language tasks, where a model tuned on Polish genuinely beats a larger general model. Oxido-style tests measure open-ended practical work, where raw capacity decides. At 4-11 billion parameters Bielik is not in the same weight class as frontier models on reasoning and code, and no published Polish benchmark contradicts that. We also found no public head-to-head of Bielik against GPT, Claude or Gemini on GSM8K, MATH or HumanEval, so we do not state the size of the gap. Anyone who states it is guessing.

One last thing to check in any offer: the 32,768-token context is confirmed only for Bielik 11B v2, and the v3 cards do not state it. 32k is enough for RAG over chunked documents and far short of the 200k-1M offered by today's frontier APIs. If your process depends on putting an entire contract into one request, that is a hard architectural limit, discussed further in the piece on an AI knowledge assistant built on RAG.

When is Bielik the wrong choice?

Honestly: in most of the processes we see on a scan, a server of your own is not the first pick. Bielik drops out in at least five situations.

• The process needs hard reasoning or code generation. Multi-step analysis, legal inference over long material, writing and fixing code: a frontier model over an API is clearly stronger here, and prompting will not close the gap.
• The volume is small. A few hundred queries a month will never repay a GPU server and its upkeep. The calculation is simple: monthly hosting and care against the same query count billed at API rates. We break down the criteria in a Polish LLM or an API model.
• You need long context. The confirmed 32k on v2 closes off scenarios where the process depends on handing over a very large body of material in one go.
• Nobody is there to maintain it. The release cadence is fast: v2 → v2.3 → v3 → the Polish-tokenizer versions in April 2026 → DFlash in June 2026 → the announced v3.1. Every swap means re-evaluation, re-quantization and regression work on prompts and evals. Somebody has to own that.
• The task is multilingual beyond Europe. v3.0-Instruct claims 32 European languages. Outside that set a general model will do better.

Bielik makes sense when the data genuinely cannot leave your infrastructure (or a regulator requires it, as DORA does with the right to audit an ICT provider), when volume is high enough that per-token billing stops making sense, or when the task is narrow and repetitive: ticket classification, field extraction from Polish documents, summarising, rewriting. The isolation levels, from cloud to air gap, are mapped in on-prem and air gap.

What a Syntalith implementation looks like

We build AI systems for companies, and that includes systems based on Bielik running on the client's own infrastructure or on EU hosting, when data constraints, GDPR posture or the cost profile call for it. We also work with API models and we say plainly that they are stronger on hard reasoning. Model choice is an engineering decision made per process. We have no named public Bielik reference deployment and we will not imply one.

Prices, net: AI automations from €3,500, AI apps and AI agents from €6,000, typical full builds €6,000–35,000, maintenance priced individually. If you want a portable document with architecture and a fixed quote before a bigger decision, the implementation specification is €1,200 net and you can take it to any vendor. If you commission the system build from us, we credit the specification fee toward the build. The full price list is on the site.

On the scan we check three things in this order: whether the process needs a model of your own at all, whether the task fits inside what an 11B model can do, and what maintenance really costs against an API at your volume. If the answer is "use the API", we will say so.

How to start

1. Book a free process scan: 30 minutes with an engineer plus a written takeaway within two business days.
2. Prepare three numbers: queries per month, what data goes into the model, and whether there is a requirement that it never leaves your environment.
3. After the call you get a recommendation: Bielik on your hardware, a model over an API, a mixed setup, or an honest "this process does not need a model".

Book a free process scan | See pricing | What an AI agent is

Related articles

• Polish language models: PLLuM and Bielik for business
• PLLuM: deploying a Polish language model in a company
• Deploying a Polish LLM: hardware, costs, GDPR
• A Polish LLM or an API model: how to choose
• On-prem and air gap: AI isolation levels
• GDPR and the DPA when deploying AI