Basal Omni Ultra Turns Gemma 31B Into a Calibrated Decision Engine
A 31B Gemma 4 fine-tune turns open language models into fast, calibrated decision engines for routing, extraction and classification in Polish and English.
Takeaways
- basal-omni-1.0-ultra is a 31B Gemma 4 fine-tune for calibrated typed decisions in Polish and English.
- Answers come as a probability distribution over up to 255 given options in one forward pass, no free text.
- Trained with LoRA on ~640,000 typed decisions, distilled from the earlier basal-1.5 model.
- Reaches 0.979 on held-out Polish decisions and 0.855 on a 10-task English decision panel.
- Supports zero-shot image inputs via Gemma 4's encoder, no audio; needs ~62 GB in bf16.
- Apache-2.0 licensed, served through the basal engine, vLLM or SGLang.
The experimental basal-omni-1.0-ultra release specializes Google’s Gemma 4 31B for bounded decision tasks in Polish and English. Given a message, document, image, or agent trace and up to 255 allowed answers, it returns a normalized probability for each answer in one forward pass. The model is the largest tier in the Basal Omni family.
For developers building classifiers, routers, or review systems, the fixed output space removes malformed responses and supports confidence-based deferral. The interface also avoids the latency and variability of an autoregressive text-generation loop.
One pass, fixed answers
Each request pairs application state with a typed question. Basal supports the following decision types:
| Type | Behavior |
|---|---|
choice |
Selects one configured option. |
noul |
Returns a binary yes-or-no decision. |
score |
Selects a score from a configured range. |
multi |
Selects multiple allowed answers through the Basal engine. |
act |
Chooses an application-defined action through the Basal engine. |
tool |
Selects from OpenAI or MCP tool definitions. Argument generation remains a separate step. |
The runtime formats the state, question, and options as a fixed chat prompt. Options receive codes from A through Z, followed by two-character codes for larger sets. The assistant turn is prefilled with the opening of a JSON response, and the runtime applies softmax only to the next-token logits assigned to valid option codes. This restricted calculation produces a probability distribution whose selected answer remains inside the configured set.
What Ultra adds
| Area | Basal Omni Ultra | Earlier tier |
|---|---|---|
| Base model | Gemma 4 31B | Bielik 11B in basal-1.5-max |
| Maximum options | 255 per question | 10 in Basal 1.5 |
| Languages | Polish and English | Polish and English |
| Image input | Supported through Gemma’s native encoder | Unavailable in Basal 1.5 |
| Audio input | Unavailable | Unavailable |
| Training data | About 640,000 typed decisions | Not reported here |
Training used soft targets distilled from Basal 1.5. Image support comes from the Gemma base model and received no additional image-specific training, so the release describes it as zero-shot multimodal support.