Qwen3.8, Bielik or PLLuM? A complete model comparison

Qwen3.8-27B is a large general-purpose model for coding, agentic tasks and work with text, images and video. Bielik-11B-v3.0-Instruct is a text model for European languages with a strong Polish focus. PLLuM-12B-chat-2512 is a Polish text model developed for public administration, dialogue and RAG. Their architectures, sizes, context windows, hardware needs and integration maturity differ substantially.

The short recommendation is:

• Choose Qwen3.8-27B for coding, tool use, visual analysis and very long inputs when the higher infrastructure cost is acceptable.
• Test Bielik-11B-v3.0-Instruct first for Polish text processing, extraction, classification, editing and a smaller local text agent.
• Test PLLuM-12B-chat-2512 first for a Polish knowledge assistant, especially over administrative documents, policies, procedures and controlled RAG.

These recommendations identify a sensible starting point. A test on the organisation's own data still decides the deployment.

The exact checkpoints in this comparison

“Bielik” and “PLLuM” are model families. A comparison without the version, size and checkpoint type is technically meaningless. This article compares three models ready for dialogue or instruction following:

1. Qwen/Qwen3.8-27B, a post-trained Qwen model with 27 billion language parameters and a vision encoder.
2. speakleash/Bielik-11B-v3.0-Instruct, the instruction-tuned Bielik 11B v3 checkpoint.
3. CYFRAGOVPL/PLLuM-12B-chat-2512, the dialogue-tuned second-generation PLLuM checkpoint released for the December 2025 line.

Older research used checkpoints including Bielik v2.3, Bielik v2.6 and PLLuM-12B-nc-chat-2412. Those results remain informative, but they do not directly describe the versions above.

Technical comparison of Qwen3.8, Bielik and PLLuM

The architectural figures come from the model cards, configuration files and technical reports. Parameter count alone does not tell us which model writes better Polish. It does reveal a meaningful difference in serving cost and in the scope of each project.

What Qwen3.8-27B is for

Qwen3.8-27B is the largest and broadest model in this comparison. It is dense, so the complete 27B language model participates in processing each token. A 27-layer vision encoder sits beside it. The model can analyse screenshots, diagrams, scanned documents, images and video, then combine that information with text and tool calls.

Its 64-layer layout is the most distinctive architectural feature. Three Gated DeltaNet linear-attention layers alternate with one full-attention layer. Full attention models token relationships effectively, but its cost rises sharply with input length. Linear layers reduce that cost. The hybrid helps support a native 262,144-token window without running full attention at every layer.

Qwen also has a thinking mode. It can be disabled for simple extraction or short answers and given more reasoning effort for debugging, planning or multistep analysis. This matters in production because longer reasoning can improve some results while increasing latency and generated-token cost.

When Qwen3.8-27B fits

• a coding agent working across a repository and terminal;
• analysis of screenshots, forms, charts or documents containing visual information;
• a workflow that combines search, code, a browser and APIs;
• long tasks where the model must maintain a plan across many steps;
• one model for Polish and international work when Polish specialisation is only one requirement.

When Qwen3.8-27B is excessive

A 27B model can waste VRAM and energy on message classification, extraction of a few fields or short Polish replies. The full context window is expensive as well. Feeding 200,000 tokens increases latency and memory use even with an architecture more efficient than a classic Transformer running full attention in every layer.

What Bielik-11B-v3.0-Instruct is for

Bielik 11B v3 is a text model developed by SpeakLeash and ACK Cyfronet AGH. It starts from the Mistral 7B v0.2 architecture. The team increased depth from 32 to 50 layers through depth up-scaling, then continued training on data from 32 European languages. The report says Polish accounted for 54.25% of training documents and English for 20.5%.

The Instruct checkpoint underwent instruction tuning, DPO and reinforcement learning. It is intended to follow commands, generate answers and run inside applications. The project's official tools add structured output, function calls and optional reasoning under vLLM. Integration requires the correct chat template and parser, so serving an endpoint is only part of the work.

The native context window is 32,768 tokens. The report describes extension to 131,072 with YaRN. An extended context still needs a task-specific evaluation because accepting a long input does not guarantee reliable retrieval and use of every fact inside it.

When Bielik 11B v3 fits

• classification of Polish letters, messages and tickets;
• field extraction and structured JSON generation;
• editing, proofreading and summarisation in Polish;
• an internal assistant working mainly in Polish;
• a smaller local text agent with tool calls;
• a multilingual environment focused on European languages.

Bielik does not process images or video. Scanned invoices and forms require OCR or a separate vision model before Bielik receives the text.

Bielik's important legal constraint

The checkpoint card says Apache 2.0, but the additional Bielik v3 terms, version 1.1 dated 31 July 2025, prohibit processing any personal data. This blocks direct use of this version on unredacted customer messages, HR documents, agreements containing people's details and many service workflows. A deployment team should check the current terms and confirm their interpretation with legal counsel. Local hosting does not remove this restriction.

What PLLuM-12B-chat-2512 is for

PLLuM is a Polish language-model programme rather than one checkpoint. Its family includes 4B, 8B, 12B and 70B models, base, instruct and chat variants, and several licences. This comparison uses the open PLLuM-12B-chat-2512 because its size is close to Bielik 11B and it is prepared for direct dialogue.

The model is based on Mistral-NeMo-Base-2407. It has 40 layers, a 5,120 hidden dimension, 32 query heads, 8 KV heads and a 131,072-token context. The project continued training on Polish and English corpora, then used manually authored instructions, RAG instructions and human-rated preference pairs. The chat variant adds alignment for dialogue, safety and general use.

RAG is PLLuM's most clearly documented practical specialisation. Its model card publishes a prompt template that tells the model to answer only from supplied documents, add [0] and [1] citations, and refuse when the context lacks an answer. A good retriever, reranker and refusal evaluation are still required, but the template is a useful starting point for a knowledge assistant.

When PLLuM-12B-chat-2512 fits

• an employee assistant grounded in policies, procedures and official decisions;
• a chatbot answering from an approved document collection;
• summarising letters and drafting administrative correspondence;
• Polish public-sector and regulated projects where model provenance and self-hosting matter;
• RAG experiments that require citations and a clear response when evidence is missing.

PLLuM-12B-chat-2512 is a text model. Its official card does not present it as a model for vision, advanced coding or autonomous agentic work. An application can surround it with tools, but the team must define the output format, parser and recovery path for malformed calls.

Which model variant to download

This distinction matters particularly in the PLLuM family:

The Bielik checkpoint in this article is an Instruct model that combines instruction following with conversational features. Qwen3.8-27B is already post-trained and supports direct responses as well as thinking mode.

GPU memory requirements

Weights alone require roughly 2 bytes per parameter in BF16, 1 byte in 8-bit and 0.5 byte in 4-bit. KV cache, buffers, activations, the vision encoder, the inference stack and concurrent requests need additional memory.

These are capacity estimates rather than speed benchmarks. Quantisation can reduce quality, especially in extraction, reasoning and code generation. Every candidate should be tested in the numeric format planned for production.

The official PLLuM site estimates about 48 GB for a 12B generator in a production RAG system without documenting quantisation. That figure provides more operational headroom than the weights alone. It does not conflict with running a 4-bit model on a smaller card; it describes a different deployment profile.

Which model handles Polish best

Bielik and PLLuM have a clear design advantage because they were deliberately continued-trained and aligned for Polish. Qwen3.8 is a more capable general model, but its card does not publish a separate matched evaluation for contemporary Polish.

The Bielik 11B v3 report compared Bielik-11B-v3.0-Instruct with an older PLLuM-12B-chat checkpoint under one methodology. Bielik scored higher on CPTUB, Polish Medical and PLCC, among other tests. The Bielik team wrote the report, and the PLLuM checkpoint predates 2512. This places Bielik high on the shortlist, but it does not settle a comparison with the current PLLuM-12B-chat-2512.

Two RAG studies on older versions further show how results depend on data:

• on real low-code platform documentation, Bielik-11B-v2.3-Instruct averaged 4.521/5 and PLLuM-12B-nc-chat averaged 4.025/5;
• on Polish Wikipedia and PolQA questions, PLLuM-12B-nc-chat received a higher rating than Bielik-11B-v2.6-Instruct.

Different checkpoints, corpora, retrievers and judges reversed the order. This is normal in real systems. A general leaderboard matters less than performance on the documents, questions and expensive failure modes present in a specific organisation.

Which model to test for each task

“First candidate” describes test order. It does not name an automatic winner.

What real Qwen tests on an RTX 3090 show

We ran Qwen3.8-27B in W4A16 on an RTX 3090 with 24 GB VRAM through a modified vLLM stack. On a long Go repository repair, Qwen Code finished in 502.14 seconds and Codex awarded the patch 100/100 under Syntalith's rubric. The same local model received 87/100 when driven through another harness and 98/100 through a Claude Code compatibility layer. Codex performed these evaluations; an external auditor did not assign the scores. On a shorter CSV import task, Qwen Code passed 4/5 hidden tests and OpenCode passed 5/5.

These runs demonstrate that Qwen3.8-27B can operate in quantised form on a single 24 GB card and that the agent harness can change the outcome as much as the model choice. They did not run Bielik or PLLuM, so they do not rank Polish models. Full details are in the Qwen3.8-27B RTX 3090 results and our agent-harness comparison.

How to run a fair company evaluation

Prepare 50–100 anonymised examples from one process. For each example, record the expected answer, permitted sources, output format, and the cases that require refusal or human review. Run the exact checkpoints with the same quantisation, or record the cost of each numeric format separately.

Measure at least:

• factual and business-rule correctness;
• completeness of required fields;
• citation accuracy and unsupported-answer rate;
• valid JSON and tool calls;
• time to first token and total response time;
• VRAM at the target context and concurrency;
• the number of cases requiring human correction;
• the effect of quantisation on the same examples.

Every generator in a RAG test must receive the same retrieved passages. Retrieval quality should be scored separately because a good generator cannot recover when the search layer supplies the wrong policy.

The final choice depends on the task

Qwen3.8-27B covers the widest range: code, tools, reasoning, images, video and long context. Its larger size makes serving more demanding. Bielik-11B-v3.0-Instruct is the most interesting compact candidate for Polish text tasks and local automation, but its current additional terms block workflows containing personal data. PLLuM-12B-chat-2512 has the clearest documented route to a Polish knowledge assistant and administrative RAG, while its family offers separate checkpoints for dialogue, instructions and further training.

A useful shortlist often contains two models. For coding and visual material, test Qwen alongside one Polish specialist. For a Polish-only RAG system, test the current Bielik and PLLuM checkpoints on the same documents. Process accuracy, operating cost and legal constraints should determine the final choice.

If you want to compare them on your own work, a free process scan can define the task, data and quality threshold. Syntalith can then build the evaluation set, run the models in an isolated environment and deploy the selected option with RAG, permissions and monitoring.

Technical 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 model and infrastructure guide
• Bielik tools for structured output, tool calling and reasoning