Qwen3.8, Bielik or PLLuM? A task-based comparison

Qwen3.8-27B, Bielik-11B-v3.0-Instruct and PLLuM-12B-chat-2512 target different jobs. Qwen combines text, image, video, long context and tool use. Bielik focuses on text in European languages, with a strong Polish focus. PLLuM develops a family for Polish dialogue, instructions and knowledge workflows.

The right choice depends on the process, data and cost of failure. A sensible testing order is:

• Qwen for code, vision, multiple tools and very long inputs;
• Bielik for Polish classification, extraction, editing and a smaller local text agent;
• PLLuM for an assistant grounded in Polish policies, procedures and documents.

This order follows published capabilities and stated use, rather than a shared test of the three versions. The organisation's own cases decide deployment.

The exact model versions

“Bielik” and “PLLuM” are families, so this article names each model version:

1. Qwen/Qwen3.8-27B, a post-trained model with 27 billion parameters in the language component and a separate vision encoder.
2. speakleash/Bielik-11B-v3.0-Instruct, the instruction-tuned Bielik 11B v3 model version.
3. CYFRAGOVPL/PLLuM-12B-chat-2512, the dialogue model version from the PLLuM 2512 line.

Earlier studies use model versions such as Bielik-11B-v2.3-Instruct, Bielik-11B-v2.6-Instruct and PLLuM-12B-nc-chat. Their figures describe those versions. We do not silently apply them to the three versions here.

Technical card

The parameter and architecture entries come from the cited cards, configurations and reports. This is not a shared benchmark. Size helps predict memory requirements; it does not decide which model writes better Polish.

Qwen3.8-27B: code, vision and long context

Qwen is the largest model here. Its dense language component has 27B parameters. A separate vision encoder accepts visual material, and the card also describes video input. The thinking mode can be disabled for simple field extraction or kept for debugging and planning.

Its layer layout combines 48 Gated DeltaNet blocks with 16 gated full-attention blocks. This hybrid supports long inputs. The native 262,144-token limit still costs memory and time. In our test, three facts at the beginning, middle and end of a synthetic record were returned at 230,085 tokens, while the run took 563.4 seconds.

Test Qwen first on tasks that involve:

• a coding agent working across a repository;
• workflows joining a terminal, browser, API and documents;
• charts, forms, screenshots and scanned material;
• long multi-step work where one session must retain a broad context.

For short classification or a few fields, the model can cost more to serve than a smaller model or parser.

Bielik-11B-v3.0-Instruct: Polish text on a smaller model

Bielik v3 is developed by SpeakLeash and ACK Cyfronet AGH. Its official card describes 11B, 33 languages and more than 20 million instructions covering 17 billion tokens. The technical report describes a Mistral 7B v0.2 base, increased depth to 50 layers and continued training on European data.

The Instruct model is intended for following requests, transforming text and generating application responses. The Bielik project publishes tools for structured output, function calls and optional reasoning. Integration needs the correct template and parser for the chosen server.

Test Bielik first on:

• classifying Polish letters, messages and tickets;
• extracting fields and producing JSON;
• Polish proofreading, editing and summaries;
• a smaller local text assistant with moderate context.

The model is documented as a text model. An invoice scan needs OCR or a separate vision model.

Bielik's licence and personal data

The card lists Apache 2.0, while the additional Bielik v3 terms add restrictions. Version 1.1 dated 31 July 2025 prohibits processing personal data. Before using this model version with customer messages, HR material or contracts, read the current terms and obtain legal review. Local hosting does not change licence wording.

PLLuM-12B-chat-2512: Polish dialogue and RAG

PLLuM is a family with multiple sizes and base, instruct and chat variants. This article covers PLLuM-12B-chat-2512, a dialogue model based on Mistral-NeMo-Base-2407. Its variant card lists 12B, 40 layers, a 5,120 hidden size and a 131,072-token context.

The chat model is documented for employee conversations and document assistance. Its card publishes a RAG prompt pattern that asks the model to answer from supplied context, add [0] and [1] references and signal when the context has no answer. A deployment still needs retrieval, access control and refusal tests.

Test PLLuM on:

• an assistant grounded in policies, procedures and official decisions;
• Polish correspondence summaries;
• RAG with required citations and an explicit evidence boundary;
• projects where keeping the model in the organisation's environment matters.

The official card describes this model as text-only. Vision, autonomous tools and advanced coding require separate components and tests.

Variants matter

A family name is insufficient for a licence, context or prompt-format decision. Record the repository identifier, file revision and chat template.

Hardware and deployment

A rough estimate for weights alone is 2 bytes per parameter in BF16, 1 byte in 8-bit and 0.5 byte in 4-bit. Context cache, buffers, activations, vision components and headroom for concurrent requests add to that budget.

These are capacity estimates rather than speed or quality benchmarks. Test quantisation on the same cases planned for production.

The weight licence is one part of a deployment review. Check the variant terms, server image source, data rights, retention, user access and tools the model may call. A managed API usually reduces maintenance. A local service can fit data-route requirements, while creating an owner for updates, monitoring and evaluation.

What we can say about Polish

Bielik and PLLuM were developed with a clear Polish-language goal. Qwen is a broader general model, and its card does not publish a matched standalone score for contemporary Polish. That leaves the quality question open for the intended process.

The Bielik v3 report compares Bielik with an earlier PLLuM model version and shows higher Bielik results on selected tests. The report was produced by the Bielik team, and that PLLuM version predates 2512. Two RAG studies on older model versions also reversed the order: Bielik-11B-v2.3-Instruct averaged 4.521/5 on low-code documentation, while PLLuM-12B-nc-chat scored higher than Bielik-11B-v2.6-Instruct in a PolQA study. Corpus, retriever, prompt, judge and model version all matter.

The fair company test uses the same Polish examples, sources, refusal rules and output format for every model version.

Test order by task

What our Qwen test shows

We ran Qwen in W4A16 on a 24 GB RTX 3090 through a modified, frozen vLLM stack. On the Go repair, Qwen Code took 502.14 seconds and Codex awarded 100/100 under Syntalith's rubric. The same local model received 87/100 through the Codex harness and 98/100 in a Claude Code compatibility experiment. Codex assigned the scores. Bielik and PLLuM did not take part.

The 100-point scale covers correctness (40), regression tests (20), compatibility (15), scope discipline (10), verification (10) and documentation (5). These scores describe the recorded runs and do not establish repeatability. This shows that the serving profile and the program guiding the model can affect the outcome. It is not a ranking of Polish models. The full Qwen results and public JSON results document those runs.

Build a company comparison

Prepare 50–100 anonymised examples from one process. Record the expected result, allowed sources, output format and cases requiring refusal or human review. Run exact model versions with the same quantisation and a comparable context limit.

Measure rule correctness, field completeness, citation accuracy, valid JSON, response time, VRAM, human correction cost and justified refusal rate. In RAG, give every generator the same retrieved passages and measure retrieval separately.

Choose Qwen, Bielik, PLLuM, a mixed architecture or an API after that test. A free process scan can define the data, quality threshold and operating cost.

Sources

• Qwen3.8-27B model card and configuration
• Bielik-11B-v3.0-Instruct model card
• Bielik 11B v3 technical report
• PLLuM-12B-chat-2512 model card
• PLLuM family technical report
• Official PLLuM site
• Bielik tools