BASAL 2.0 + BASAL-OMNI
basal-2.0 reads text. basal-omni also accepts images and recordings. You define the questions and allowed answers. The model chooses between them β locally, with a confidence estimate.
System-1 models Β· a classifier defined with every request
The same dynamic classifier concept. Choose text models for fast automation, or multimodal models when the answer is in an image or recording.
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.
Model downloads will be available after launch.
NVIDIA GPU Β· Mac / MLX Β· Ollama / llama.cpp
Run a text modelDecisions from what you see and hear, too.
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.
Model downloads will be available after launch.
NVIDIA GPU Β· bf16 / eager Β· approx. 16 GB mini, 24 GB max
Run a multimodal modelbasal-omni is our first release with media: images and audio use the base model's capabilities without additional multimodal training. Calibration comes from text. Check results and limitations β
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.
Text for basal-2.0. Text, images or audio for basal-omni.
βThe shipment arrived damagedβWhich team? How urgent? Which topics? Several questions about one context in the same call.
Sales Β· Complaints Β· LogisticsGet the result and probabilities. Your application takes the next step or hands the case to a person.
Complaints β support queuechoice one optionnoul yes / noscore ratingmulti multiple labelsact action with costsAPI capabilities βWerdykt is our hidden benchmark: 10 categories in Polish and English. We show five new basal models and three reference points β a cloud API model, an open model and Jev.
basal-2.0 focuses on speed and hardware flexibility. basal-omni also handles text; its max variant has the highest accuracy in the current basal family.
These are text results. Image and audio quality are shown in the other tab.
See all systems βTime and cost: shorter bars are better. Each scale runs from zero to the maximum in this comparison.
Werdykt PL+EN: basal-omni-max 82.28%, omni-mini 76.88%. basal-2.0: max 78.40%, 4.5B 72.86%, mini 61.84%. Results and measurement conditions β
Werdykt measures text decisions. Full data and comparison conditions are on the benchmark page.
Five small Polish probes, from document images to speaker intent. Here, the input is an actual image or audio recording.
Max performs better 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.
Full multimodal results βAccuracy from media Β· 0β100% scale Β· higher is better
Find all five probes in the multimodal table β.
omni-probe-pl Β· 160 + 150 image questions, 140 + 100 + 45 audio questions. These are separate probes, not Werdykt results.
The previous answer changes the question, operation type and next set of options.
See how it works β
The document image goes into the model. Questions turn its content into decisions.
Watch document analysis β
A recording, questions about its meaning and the selected follow-up path.
Listen and see the result βThe recordings show earlier model versions. Each case lists the version and actual measurements. Videos are in Polish. Open the use case catalog β
Run basal-2.0 for text or basal-omni for images and audio. Both accept questions through POST /v1/systemone. Switch models, keep the same integration.
Models and engine launch: Wednesday, 14 October at 09:00 (Warsaw).
Full documentation# 1. Create an environment (requires uv)
uv venv --python 3.12 ~/basal-env
source ~/basal-env/bin/activate
# 2. Install PyTorch and the engine
uv pip install torch==2.11.0 \
--index-url https://download.pytorch.org/whl/cu128
uv pip install "basal[fp8] @ https://github.com/rkinas/basal/archive/refs/tags/v2.0.0.tar.gz"
# 3. Start the model (bf16; FP8 port: Remek/basal-2.0-4.5B-FP8 --mode fp8)
basal-serve --model Remek/basal-2.0-4.5B \
--mode fast --port 8000Install 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.
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.
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.
Images and audio recordings up to 30 seconds, including alongside text in one request. This release does not accept video files directly. Your application can extract frames and the audio track first.
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.
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 β
basal-2.0 and basal-omni: models, results, calibration and practical API examples. All the details in one article.
Building with basal? Share your use case β
See who uses basal β