SpeakLeash and ACK Cyfronet AGH announced a three-day Bielik model rollout on 10 November 2026: Bielik 11B v3.1, its DFlash accelerator and four Bielik-One decision models. Their Hugging Face repositories now list weight files, giving researchers separate models for text generation, faster inference and structured classification. The announced launch schedule and the available artifacts are distinct evidence.
Bielik 3.1 rollout: dates and available weights
The November 10 announcement scheduled the main model for November 10, DFlash for November 11 and Bielik-One for November 12. The repository inventories establish that artifacts are now listed; their creation or modification timestamps do not establish when public access began.
| Model | Announced launch | Artifacts listed on Hugging Face |
|---|---|---|
| Bielik 11B v3.1 Instruct | 10 November 2026 | Sharded Safetensors weights; separate GGUF and other quantized repositories. |
| Bielik 11B v3.1 DFlash | 11 November 2026 | A model.safetensors weight file in the draft-model repository. |
| Bielik-One family | 12 November 2026 | Weights for 4B-r1, 7B-v3.0-r1, 11B-v3.0-r1 and 11B-v3.1-r1, alongside adapter repositories. |
The main Instruct repository identifies an 11-billion-parameter model under Apache 2.0 and requires users to accept access conditions for its files. An open weights model can therefore still have a repository access step.
Tool use and acceleration have different roles
Cyfronet describes the main release as combining function calling, structured output and reasoning for tasks involving multiple steps. It reports gains of 16–45 percentage points over Bielik 3.0 on τ²-bench, depending on the scenario. These are developer-reported results from customer-service simulations covering airlines, retail and telecommunications; they do not establish performance on university research administration.
The DFlash documentation identifies it as a companion to Bielik-11B-v3.1-Instruct, requiring an inference server that supports its speculative decoding. The helper proposes response fragments for the main model to verify. SpeakLeash says this accelerates generation without reducing response quality; that statement is a developer claim, rather than an independent performance measurement.
For university computing teams, the distinction is practical: DFlash supplies an acceleration component, while Instruct supplies the underlying text-and-tool model. The Instruct model card also documents specific tool-call and reasoning parsers, warns that outputs can be factually incorrect and says the model lacks moderation mechanisms.
Bielik-One brings structured decisions to the family
The Bielik-One 11B v3.1 documentation provides full merged weights and an adapter alternative. Its decision interface supports choices, binary questions, scores and multiple labels, including probabilities for the available options. This is the part of the rollout aimed at returning a defined decision rather than an extended answer.
For researchers and research offices, our interpretation is that the release creates distinct candidates for document-routing experiments and text-based workflows, with separate evaluation needs. The One card supplies a concrete caution for a reproducible AI experiment: its probability calibration was fitted to bfloat16 outputs and does not cover FP8 serving. A confidence value from one serving configuration therefore should not automatically be treated as equivalent to another.
Sources
- Bielik AI: launch announcement and schedule, 10 November 2026. Checked 2026-11-14.
- ACK Cyfronet AGH: Premiera Bielika 3.1., 10 November 2026. Checked 2026-11-14.
- SpeakLeash: Bielik-11B-v3.1-Instruct model card and access conditions. Checked 2026-11-14.
- SpeakLeash: Bielik-11B-v3.1-DFlash model card. Checked 2026-11-14.
- SpeakLeash: Bielik-One-11B-v3.1-r1 model card and calibration. Checked 2026-11-14.
- Hugging Face: SpeakLeash public model file inventories. Checked 2026-11-14.
