basal-omni-1.0-ultra (experimental)

basal-omni-1.0-ultra (31B) is the largest model of basal-omni, basal's line on Google Gemma 4. It is an experimental release: it is trained on text decisions only, and it also answers the same questions about images, read by Gemma 4's own image encoder without media training (zero-shot; media training comes in the next releases). The 31B base has no audio encoder, so this tier takes no audio. It answers the same typed questions as basal-1.5, over the same API.

What it is. Inspired by System 1 (fast, intuitive) decision models such as Jev. Instead of writing an answer, the model reads a state (a message, a document, retrieved passages, an agent trace, a web page as JSON). It then answers a typed question about it (choice, yes/no noul, score, and in the basal engine multi, act and tool; a tool question over OpenAI or MCP tool definitions picks the tool, not its arguments). The answer is a calibrated probability for each allowed answer, in one forward pass, without generating text. The answer never falls outside the options you give, and the probability says how sure the model is.

What is new against basal-1.5-max.

• A Gemma 4 base (31B) instead of Bielik (11B).
• Up to 255 options per question (basal-1.5: 10).
• One option order by default (basal-1.5 averaged two).
• Images in the same request ("media", served with --media on), zero-shot; no audio.
• Trained on about 640,000 Polish and English typed decisions, with soft targets from basal-1.5.

Results

Served with the basal engine: basal-omni-1.0-ultra on one H200 and basal-omni-1.0-max on one H100, both with --mode eager, bf16, one option order; basal-1.5-max as released (--mode fast, both option orders). Same items and the same clients for all three. Measured with basal 2.0; basal 2.1 gives the same top answer on 23,278 of 23,278 served-check rows.

• Against basal-omni-1.0-max: ahead on PL decisions (+2.5 points, 95% interval +2.0 to +3.1), the English corpus test of 1,479 items (0.751 vs 0.724, +0.6 to +4.7) and the English external panel (+6.9 points over its items; ahead on 9 of 10 tasks: WinoGrande 0.863 vs 0.710, BIG-Bench Hard 0.883 vs 0.809, JudgeBench 0.700 vs 0.600; behind on FinancialPhraseBank, 0.947 vs 0.967). Level on Polish general knowledge (+1.3, 0.0 to +2.7) and PL-Real (−0.7), whose intervals include zero.
• Against basal-1.5-max: PL decisions +3.9 points (+3.3 to +4.5), Polish general knowledge +2.6 (+1.1 to +4.1), English corpus test 0.751 vs 0.659, English panel +11.8; level on PL-Real (+0.3).
• ¹ Coverage: per question type, at 0.47% observed error (max: 47.4% at 0.47%). A type without a certified threshold is deferred: at the 1% target, score questions go to a person. basal-1.5-max's 62.2% used one threshold pooled over all types; under that rule this model reaches 77.5% at 0.75% error.
• Werdykt v1: 0.879 (95% interval 0.871–0.887; Polish 0.881, English 0.877), 5.6 points above basal-omni-1.0-max (0.823) and 10.6 above basal-1.5-max, served with --deterministic on one H200 (one request per forward). Ahead of max in all 10 categories; weakest categories: rules (0.550; compute exact thresholds and dates in code), contracts (0.822) and long documents (0.854).

Accuracy by document length

Accuracy of basal-omni-1.0-ultra on Werdykt (5,000 items, Polish and English; a hidden set, so only these totals are shown) by the length of the state, in characters (a token is about 2 characters; 1,000 characters is about half a page). 87.9% overall. 95% Wilson intervals in brackets.

Short states are mostly rules and abstention items, which are harder for every basal model, so read the rows as accuracy per bucket, not as the effect of length alone. The long-document family keeps the task and varies only the length:

Run 2026-10-15 with basal-serve --mode eager on one NVIDIA H200 (bf16). The intervals are about the items, not about run-to-run noise.

By prompt length in tokens (held-out tests)

The held-out tests of the Results table by the length of the whole prompt as served (state, question and options; this model's tokenizer), with the Results table's predictor. Accuracy (items):

Quick start

1. Install into a fresh uv environment:

uv venv --python 3.12 --allow-existing ~/basal-env && source ~/basal-env/bin/activate
uv pip install torch==2.11.0 torchvision==0.26.0 --index-url https://download.pytorch.org/whl/cu128
uv pip install "basal[media] @ https://github.com/rkinas/basal/archive/refs/tags/v2.5.3.tar.gz"

2. Start the server. It prints basal: model ... ready:

basal-serve --model Remek/basal-omni-1.0-ultra --mode eager --media on --port 8000

The eager mode runs the model in PyTorch, bf16 (about 62 GB of GPU memory for the weights; on one 80 GB GPU, states up to about 10,000 tokens, see Limitations). For text only, the FP8 checkpoint basal-omni-1.0-ultra-FP8 runs on --mode vllm. Since basal 2.3, text-only requests can also use the fast mode (shared prefix, CUDA graphs): the default without --mode on Hopper GPUs with the flash_attn_3 package and on Blackwell workstation GPUs (2.3.1 or later), eager on other GPUs. Details: OMNI.md.

3. Ask questions, exactly as with basal-1.5:

curl -s localhost:8000/v1/systemone -H 'content-type: application/json' -d '{
  "state": "Klient: od wczoraj nie mogę zalogować się do bankowości internetowej, system pokazuje błąd hasła.",
  "questions": {
    "dept":   {"type": "choice", "instructions": "Do którego działu skierować zgłoszenie?",
               "criteria": {"cards": "Reklamacje kart", "online": "Wsparcie bankowości elektronicznej", "loans": "Kredyty"}},
    "urgent": {"type": "noul", "instructions": "Czy klient nie może korzystać z usługi?"}}}'

Full API (multi, act, tool, option keys, max_error): API.

Images

With --media on, a request may carry images with its state. They are placed where the state text goes, before it, in the order given; the state text may be empty.

IMG=$(base64 < paragon.jpg | tr -d '\n')
curl -s localhost:8000/v1/systemone -H 'content-type: application/json' -d '{
  "state": "", "lang": "pl",
  "media": [{"type": "image", "data": "'"$IMG"'"}],
  "questions": {
    "card": {"type": "noul", "instructions": "Czy zapłacono kartą?"},
    "vat":  {"type": "choice", "instructions": "Jaka jest najwyższa stawka VAT na paragonie?",
             "criteria": {"23": "23%", "8": "8%", "5": "5%", "0": "0% lub zwolniony"}}}}'

• "media": a list of {"type": "image", "data": "<base64>"} (a data: URL works too), any format Pillow reads. Audio is refused with HTTP 422 (this model does not take audio (it takes: image)): the 31B base has no audio encoder.
• "detail": "default" (280 image tokens) or "high" (1,120; small print, slower).
• "lang": "pl" or "en", the prompt language when the state text does not show it (a state that is only media).
• Limits: 8 items and 20 MB each per request, 50 MP per image.
• Every question of a request sees all of its images. For the steadiest answers, send each question with only the images it is about.
• Answers carry "calibration": "text": the temperatures and the certified thresholds were fitted on text; validate the thresholds on your own media before automating with them.

On omni-probe-pl, our small Polish probe (each item asked from the image and from its text):

• Rendered documents are read almost as well as their text (4 points behind) and 12.5 points better than by the untrained base (95% interval +5.6 to +19.4).
• Photos add 19 points over the question alone. The text training did not improve photos: 3.3 points below the untrained base, within the probe's noise (95% interval −9.3 to +2.7).

Speed

One question per request, one option order, on one H200 (--mode eager, bf16, --media on):

• basal-omni-1.0-max on one H100, same test: 66 ms p50 and 19.0 decisions/s; basal-1.5-max in its fast engine: 27.3 ms p50 and 33.4 decisions/s.
• --orders 2 costs about twice as much per question.
• Long states are prefilled in 2,048-token chunks; the longest state a GPU serves is under Limitations.

Quantised modes (basal >= 2.5.0)

basal 2.5.0 serves this model in three quantised modes of its CUDA graph engine. Each mode loads its own calibration file from this repository, fitted on that mode's readouts; without it, int8 and nvfp4 refuse the model. The files are fitted for one option order (the default). Details: QUANTIZATION.md.

Files

• model-*.safetensors, tokenizer, chat_template.jinja, processor_config.json: the model.
  • Gemma 4 architecture, 31B, 262k vocabulary.
  • Gemma's image encoder is included and unchanged (this release is trained on text); the engine reads images through it with --media on (processor_config.json). The 31B base has no audio encoder.
• CALIBRATION.json: per-type temperatures and certified confidence thresholds for 1% / 5% error, fitted on the served engine (--mode eager, bf16, one option order).
• CALIBRATION.o2.json: the same for --orders 2 (both option orders averaged).
• basal.json: prompt format 2 and the readout protocol.

How it works

The model receives a fixed chat prompt with the state, the question and coded options (A–Z, then two-character codes, up to 255). The assistant turn is prefilled with {"answer": ", and the decision is the softmax over the next-token logits of the option codes only: one forward pass, no text generation. The server applies the calibrated temperature of the question type from CALIBRATION.json.

Use the confidence. CALIBRATION.json stores thresholds for a target error of 1% or 5% among accepted decisions. They were chosen on calibration data before testing; on the test data they accept 56.8% / 94.2% of decisions at 0.47% / 3.8% observed error (at the 1% target score questions have no certified threshold and are deferred). Accept decisions above the threshold automatically and route the rest to a person.

Training

• Base and method: fine-tuned from google/gemma-4-31B-it with LoRA (rank 32) on the language-model layers, merged into the weights. The image encoder was not trained.
• Data: about 640,000 typed decisions in Polish and English: rules and deadlines, real documents, long documents, retrieval decisions, contracts, routing, ordinal scores, judging answers, agent and tool decisions, and "cannot be determined" answers.
• Distillation: basal-1.5 provides soft targets on the decisions it answers correctly.
• Schedule: 10,000 of the 12,741 planned steps of 64 decisions (640,000 of the basal-1.7 mix's 815,386, without its 1,931 items longer than 6,144 tokens) at peak learning rate 3e-5 with cosine decay; stopped at step 10,000 because the dev panel had stopped improving (0.908, 0.909 and 0.909 at steps 8,000, 9,000 and 10,000), then the LoRA was merged into the weights in float64 and rounded once to bf16.

Limitations

• Experimental. It is the first model of a new base; validate it on your own documents before relying on it.
• Images are zero-shot. Gemma 4's encoder reads them, but this model was trained on text decisions only; image answers are weaker than text where the image itself is hard (phone photos), see "Images". Their calibration is the text calibration. No audio: the 31B base has no audio encoder.
• Size and speed: 31B parameters, about 62 GB in bf16, about 2× slower than basal-omni-1.0-max (see Speed). With engine 2.3.1 it answers states of up to 82,000 tokens on one 80 GB H100, in --mode fast (the state is prefilled in chunks; 82k tokens with five questions: 46 s) and in --mode eager (99 s), also while short requests are being served. A request that still runs out of GPU memory gets HTTP 503 (retry later or elsewhere), the server frees its memory and keeps serving. Engine 2.2 and earlier: about 10,000 tokens on an 80 GB H100 (8,000 with --orders 2) and 32,000 on a 141 GB H200, with HTTP 422 beyond that.
• Exact arithmetic and thresholds are a weak spot of one-pass decisions: compute them in code.
• Option order: one order is asked by default; --orders 2 reduces sensitivity to option order at twice the cost.
• tool with no fitting tool: ultra usually answers "no suitable tool" (4 of 6 out-of-scope requests), but is sometimes unsure between that and "ask the user" (1 of 6 near 0.5).
• Review: decisions with serious consequences for people should be reviewed by a person.

Citation

@software{kinas2026basal20,
  title   = {basal-2.0: typed-decision models and inference engine for Polish and English},
  author  = {Kinas, Remigiusz},
  year    = {2026},
  version = {2.0.0},
  url     = {https://github.com/rkinas/basal}
}

@techreport{kinas2026basal,
  title       = {basal-1.0: Reliable, Highly Optimized Typed Decisions for Polish},
  author      = {Kinas, Remigiusz},
  institution = {ai5},
  year        = {2026},
  type        = {Technical report},
  doi         = {10.5281/zenodo.23022986},
  url         = {https://doi.org/10.5281/zenodo.23022986}
}

License and attribution

Apache-2.0. Fine-tuned from google/gemma-4-31B-it (Apache-2.0).