Not tested by us yet. Bielik-One is a weights release with no hosted API, and we have not run it. Everything below comes from the Bielik launch post and the Hugging Face model cards for Bielik-One-11B-v3.1-r1 and Bielik-One-4B-r1. We'll update this post when we can serve the weights and run them against other decision models on our own items.
Bielik-One is the first decision-model family from SpeakLeash, the Polish open-science project behind the Bielik LLMs. It launched on 12 November 2026 as four checkpoints, from a 4B model that needs about 12 GB of GPU memory to an 11B model that needs about 30 GB, all labelled Apache-2.0 on their cards. If you want a decision model you can run yourself, with Polish-trained roots, it is the most flexible set of sizes we have seen so far. If you need a number for accuracy, there is almost nothing to go on yet.
| Bielik-One-4B-r1 | Bielik-One-11B-v3.1-r1 | |
|---|---|---|
| Maker | SpeakLeash | SpeakLeash |
| Base model | Qwen3.5-4B | Bielik-11B-v3.1-Instruct |
| Listed size | about 5B (BF16 and F32 tensors) | 11B (BF16) |
| GPU memory, bf16 | about 12 GB | about 30 GB |
| License | apache-2.0 | apache-2.0 |
| Context in default profile | 32,768 | 32,768 |
| Reported accuracy | 0.69 on Typed Decisions (400 cases, 2,000 decisions) | none on the card |
| Hosted price | none | none |
What a decision model does
A chat model writes tokens that you then have to parse. A decision model returns probabilities directly. You send a state (the thing to judge) and a set of questions, and each question comes back with an answer and its probabilities. The cards list four question types:
- choice: one option from a set of keys, with a probability per option and a confidence.
- noul: a yes/no question, returned as the probability of yes.
- score: ordered levels, returned as an expected level plus probabilities.
- labels: several independent options, each with its own probability, and a
selectedlist of those at 0.5 or above.
The first three match the types in Kanana-2-v-decide and Salesforce's Koa Decisions. labels is the multi-select addition. The launch post gives a plain example: for each email in an inbox, the model outputs a probability per category, and you set thresholds so clear cases are automated and uncertain ones go to a person. Other use cases it lists are request routing, content moderation, screening applications and documents, checking procedural compliance, sentiment analysis and evaluating answers from other AI systems.
The family
The launch post says the models come in different sizes but does not name them. The SpeakLeash Hugging Face page lists four base checkpoints, each with a matching Adapter repository:
- Bielik-One-4B-r1
- Bielik-One-7B-v3.0-r1
- Bielik-One-11B-v3.0-r1
- Bielik-One-11B-v3.1-r1
We read the 4B and 11B v3.1 cards in full. We only saw the 7B and 11B v3.0 entries in the list, so treat their specs as unconfirmed.
The Adapter option is a practical detail. The 11B card says its adapter adds about 1.0 GB to an existing Bielik-11B-v3.1-Instruct deployment, and the 4B adapter adds about 0.7 GB. If you already host the Instruct model, you can add decisions without a second full copy of the weights. The merged checkpoints are built for stock vLLM 0.32.0, with no plugin and no runtime LoRA, installed through the bielik-one package from the SpeakLeash GitHub repository.
Two calibration notes from the cards matter if you plan to use the probabilities. The 11B card gives per-type temperatures (2.107 for choice, 1.899 for binary, 2.732 for score, 1.000 for labels) and says calibration does not change the top answer or a 0.5 threshold. It also says --fp8 roughly halves weight memory but is not calibrated and can change some decisions near the boundary. Both cards say the same about FP8.
How it compares on paper
| Bielik-One | Gemini Decision 1 | Perplexity Decider v1.3 | TypeSafe Jev | |
|---|---|---|---|---|
| Access | Open weights, no hosted API | Hosted | Weights, gated repo | Hosted, early access |
| Sizes | 4B to 11B | Not stated | 26B | Not stated |
| Question types | choice, noul, score, labels | See our write-up | choice, noul, score, locate | See vendor post |
Our write-ups of Gemini Decision 1 and Perplexity Decider v1.3 cover those two in more detail. No vendor in this table publishes a comparison against the others that we could verify, and the only accuracy figure for Bielik-One is the 4B card's 0.69 on its own Typed Decisions split. The card does not say how a general chat model scores on the same split, so 0.69 has no baseline.
What is missing
- Benchmarks. The 11B card refers to reported results but gives no figures. The launch post has none either. The 4B card shows an accuracy of 0.69, a KL-from-gold of 0.19 and a Brier score of 0.1, and says it ran on one NVIDIA A100. It also hints at one more metric that was not visible when we read it.
- Polish results. Nothing we read shows how the models do on Polish text compared with English, even though that is the obvious reason to pick them.
- Release dates per checkpoint. The cards give a version and the year 2026, not a day.
- License confirmation for the launch. The post says the weights go out on Hugging Face under open-science terms through SpeakLeash, without naming a license. The two cards we read say apache-2.0, which is what we use here.
Who should care
- Polish and Central European teams that want routing or moderation to run inside their own network, on a model family from a local project.
- Anyone with a single 16 GB to 24 GB GPU who wants a decision model: the 4B at about 12 GB is the only size that fits comfortably there.
- Not yet for teams that need proof. With one self-reported accuracy number and no head-to-head, you would have to run your own evaluation before trusting it for production routing.
What we will test when we can run it
One set of 30 to 60 items, with control items that have obvious answers and a mirrored wording for each yes/no question. We will put the 4B and 11B models next to Jev, Gemini Decision 1 and a cheap and a mid-priced chat model acting as classifiers, with a few Polish items in the mix, and report accuracy, controls passed, median latency and cost per 1,000 decisions. Until then, our Kanana-2-v-decide write-up is the closest reference for how an open decision model's claims should be read.