Polish decision model basal-Omni ranks #1 open-weight on JevBench

Polish developer basal released basal-omni-1.0, a multimodal decision model it says ranks #1 among open-weight models on JevBench, with weights published on Hugging Face in bf16 and an ultra 31B variant also offered in FP8 for vLLM from Saturday 17 October at 09:00. basal-omni accepts text, images and audio, letting users define questions and allowed answers in each request rather than training a separate classifier, and returns a choice with a confidence estimate; the text-only basal-2.0 is built on the Polish Bielik foundation while basal-omni uses Gemma 4. The company said basal-omni is its first release with media, using the base model's capabilities without additional multimodal training, and that calibration comes from text, with ultra reading text and images but not audio.

Decisions from text. Now also from images and audio. basal-2.0 reads text. basal-omni also accepts images and recordings. You define the questions and allowed answers. The model chooses between them — locally, with a confidence estimate. ONE DECISION INTERFACEYour questions. Structured answers. CategoryUrgencyNext action System-1 models · a classifier defined with every request 01 / CHOOSE YOUR MODEL Two families. A shared language of decisions. The same dynamic classifier concept. Choose text models for fast automation, or multimodal models when the answer is in an image or recording. TEXTMain release basal-2.0 Fast decisions in your workflow. Messages, documents, RAG and the agent's next tool. Up to 255 options per question, multiple questions about shared context. Polish foundation: Bielik. A document scan, a product photo, a recorded customer request. Send the media and ask about its content, without a separate OCR or transcription step. Foundation: Gemma 4. Weights on Hugging Face: basal-omni-1.0 collection · bf16. New from Saturday 17 October, 09:00: ultra 31B, also FP8 for vLLM and the experimental max 1.1, trained on the tasks of engine 2.1. basal-omni is our first release with media: images and audio use the base model's capabilities without additional multimodal training. Calibration comes from text. ultra reads text and images, not audio. Check results and limitations → 02 / HOW BASAL WORKS The classifier changes with your question. You don't train a separate model for every list of categories. You describe them in the request. The next question can address an entirely different process. 01 Provide the data Text for basal-2.0. Text, images or audio for basal-omni. “The shipment arrived damaged”02 Ask your questions Which team? How urgent? Which topics? Several questions about one context in the same call. Sales · Complaints · Logistics03 Use the decision Get the result and probabilities. Your application takes the next step or hands the case to a person. Complaints → support queue choice one optionnoul yes / noscore ratingmulti multiple labelsact action with coststool agent toolclassifyverifycompressextract tasks · basal 2.1All schemas ↗ IN PRACTICE Where can you use basal? Examples in brief. Wherever the next step of a process depends on what a document, message or recording says. You describe the question and the allowed answers; the model returns a decision with a confidence level. Every question schema and the basal 2.1 tasks. Explained step by step. Every question type with a full JSON example: from choice, noul and score through multi, act and tool to SOAM, calibration and the classify, verify, compress and extract tasks. Werdykt to nasz ukryty benchmark: 10 kategorii po polsku i angielsku. Pokazujemy sześć modeli basal oraz punkty odniesienia — model API, otwarty model, Jev 1.14, Liquid d1-ultra i Respan Span-02 w dwóch ustawieniach: z rozumowaniem i bez niego jeden przebieg, jak basal . 5 000decisions · the same protocol basal-2.0 focuses on speed and hardware flexibility. basal-omni also handles text; its ultra 31B has the highest accuracy in the basal family. These are text results. Image and audio quality are shown in the other tab. Werdykt measures text decisions. Full data and comparison conditions are on the benchmark page. What does the model read from images and recordings? Five small Polish probes, from document images to speaker intent. Here, the input is an actual image or audio recording. mini · max · ultraon the same questions; ultra images only Ultra reads images best but takes no audio. Max does better than mini on document images. Mini handles recordings better. Model size is no substitute for testing your task. Experimental release. Small sample, text-based calibration. The full table compares media with text and the base model. Install uv and a driver supporting CUDA 12.8 B300 and HP ZGX Nano Nvidia DGX Spark : cu130 . The first start compiles the model for a few minutes. All GPUs: hardware guide. GOOD TO KNOW Which basal fits your project? Text or multimodal? Choose basal-2.0 for fast text decisions on GPU, Mac or through Ollama. Choose basal-omni for image or audio input. Omni also answers text questions, but currently runs only on NVIDIA GPUs in eager mode. Does the model write answers like a chatbot? By default, it returns a typed decision and probabilities for the allowed answers. Use a generative model to write an email or summary; basal can select a path, assess the input or check the result. The release article describes the experimental thinking mode in basal-2.0 4.5B. What is new in basal 2.1? Engine v2.1.0 adds four experimental tasks on the same API: classify labels with a rubric each , verify whether statements follow from the text , compress the parts a question needs and extract JSON schema fields copied exactly from the document . The model does not write text: a task is built from its decisions. On Saturday 17 October basal-omni-1.0-ultra 31B and the experimental task model basal-omni-1.1-max arrive. See the tasks → Which media does basal-omni support? Images and audio recordings up to 30 seconds, including alongside text in one request ultra: images only . This release does not accept video files directly. Your application can extract frames and the audio track first. Does data have to leave the company? No. After downloading the weights, you run the model on your own hardware and control the application and its connections. Inference itself needs no external API. What does the confidence level mean? It is a calibrated model estimate, not a guarantee of correctness. Choose a threshold using examples from your own process. basal-omni uses calibration fitted on text; it needs separate validation on images and recordings. Explore calibration measurements → LAUNCH / WEDNESDAY, 14 OCTOBER Explore the full release. basal-2.0 and basal-omni: models, results, calibration and practical API examples. All the details in one article.