Basal-omni-1.0-ultra brings zero-shot image decisions to Gemma 4

On October 18, 2026, developer Remek released basal-omni-1.0-ultra, a specialized open-weight model hosted on Hugging Face. Built on Google’s Gemma 4 architecture, this 31-billion parameter system is designed for high-stakes classification tasks rather than generative chat. It introduces zero-shot image understanding to the basal line, allowing engineers to route documents and queries based on visual inputs without additional media training.

What happened

This release marks the largest model in the basal-omni series, scaling up from the previous 11-billion parameter Bielik base to Google’s Gemma 4-31B. Unlike standard large language models that generate verbose text responses, basal-omni-1.0-ultra operates as a "System 1" decision engine. It reads a defined state—such as a customer message, a retrieved document, or an agent trace—and outputs a typed decision. The model supports choices, yes/no answers, scores, and tool selection, providing a calibrated probability for each option in a single forward pass.

A significant advancement in this version is the inclusion of image processing. By leveraging Gemma 4’s native image encoder, the model can analyze images alongside text states in a zero-shot manner. This means it was not explicitly trained on image-decision pairs, yet it can interpret visual data like receipts or forms. However, the model does not support audio input, as the underlying 31B base lacks an audio encoder. The release is labeled experimental, urging developers to validate performance on their specific datasets before full deployment.

The training process involved fine-tuning the Gemma 4 base using LoRA on approximately 640,000 typed decisions in Polish and English. These decisions covered diverse scenarios including contract analysis, rule checking, and agent tool selection. The model uses soft targets distillated from the earlier basal-1.5 model to improve accuracy. Training stopped after 10,000 steps when development metrics plateaued, and the weights were merged and rounded to bf16 precision for efficient inference.

How it works

The core mechanism of basal-omni-1.0-ultra differs fundamentally from generative AI. Instead of predicting the next word in a sequence, the model predicts the next token from a constrained set of option codes (A–Z and two-character combinations). When a request arrives, the server prefills the assistant’s response with a specific JSON structure. The model then performs a single forward pass, calculating softmax probabilities only for the allowed answer tokens. This approach ensures the output never falls outside the provided options and provides a direct measure of confidence.

For image inputs, the system accepts base64-encoded images or data URLs within the request payload. The Gemma 4 image encoder processes these visuals into tokens, which are then treated as part of the state context. While the encoder itself remains unchanged from the base model, the decision head learns to interpret these visual tokens through its text-based training. The server applies temperature calibration from a provided JSON file to adjust confidence thresholds, allowing users to automate decisions only when the model’s certainty exceeds a predefined error rate, such as 1% or 5%.

Key details

• Model Architecture: Based on Google Gemma 4-31B with 262k vocabulary; uses LoRA fine-tuning merged into bf16 weights.
• Input Capacity: Supports up to 255 options per question and states up to 10,000 tokens on an 80 GB GPU.
• Media Support: Handles images via Gemma 4’s encoder in zero-shot mode; explicitly rejects audio inputs with HTTP 422 errors.
• Performance Metrics: Achieves 0.879 macro score on Werdykt v1 benchmark, outperforming the 12B max variant by 5.6 points.
• Inference Speed: Runs at 80 ms p50 latency per question on an H200 GPU in eager mode, supporting about 8.6 decisions per second under load.
• Calibration: Includes certified confidence thresholds for 1% and 5% error rates, enabling safe automation by deferring low-confidence items to humans.

Why it matters

For software engineers building classification layers, this model offers a deterministic alternative to probabilistic text generation. Traditional LLMs often hallucinate or drift from strict formatting requirements, requiring complex post-processing pipelines to extract structured data. Basal-omni-1.0-ultra eliminates this overhead by constraining outputs to predefined types. This reliability is crucial for applications like customer support routing, compliance checking, or financial document processing, where incorrect categorization can have serious operational consequences.

The integration of zero-shot image understanding expands the model’s utility without the cost of massive multimodal training sets. Developers can now build systems that read scanned invoices, identify payment methods from receipt photos, or classify handwritten forms using the same API endpoint used for text. While the accuracy on difficult real-world photos lags behind text-based analysis, the performance on rendered documents is competitive. This allows teams to unify their text and vision decision logic into a single, manageable service.

What you can do

• Install the model in a fresh Python environment using uv and PyTorch 2.11.0 with CUDA 12.8 support.
• Start the server with the --media on flag to enable image processing capabilities for your requests.
• Test the model with your own document images using the provided curl examples to baseline accuracy against your current workflow.
• Review the CALIBRATION.json file to set appropriate confidence thresholds for your specific error tolerance levels.
• Use the --orders 2 flag if option order sensitivity is a concern, accepting the trade-off of doubled computational cost.
• Validate the model on a held-out set of your real-world data before automating any high-stakes decisions.