basal-omni-1.0-max (experimental)

basal-omni-1.0-max (12B) is the first 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 and audio, read by Gemma 4's own encoders without media training (zero-shot; media training comes in the next releases). 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 (12B) 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 and audio in the same request ("media", served with --media on), zero-shot.
• Trained on about 815,000 Polish and English typed decisions, with soft targets from basal-1.5.

Results

Served with the basal engine on one H100: basal-omni-1.0-max 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 both. Measured with basal 2.0; basal 2.1 gives the same top answer on 23,276 of 23,278 served-check rows.

• Werdykt v1: 0.823 (95% interval 0.813–0.832; Polish 0.824, English 0.821), 5.0 points above basal-1.5-max, served with --deterministic (one request per forward; at most 0.1 point from the default mode on our tests). Weakest categories: rules (0.418; compute exact thresholds and dates in code), long documents (0.742) and contracts (0.762).
• English: it is ahead on the English external panel (+5.2 points; BIG-Bench Hard 0.809 vs 0.644, TabFact 0.873 vs 0.750; behind on ContractNLI, 0.720 vs 0.740), and on the English corpus test of 1,479 items (0.724 vs 0.659).
• Polish: it is ahead on the PL decisions test (+1.4 points, 95% interval +0.8 to +2.0), and level on Polish general knowledge (+1.3) and PL-Real (+1.0), whose intervals include zero.
• ¹ Coverage: per question type, at 0.47% observed error. 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 65.1% at 1.02% error.

Accuracy by document length

Accuracy of basal-omni-1.0-max 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). 82.3% 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 2026-10-11 with basal-serve --mode eager on one NVIDIA H100 80GB HBM3 (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-max --mode eager --media on --port 8000

The eager mode runs the model in PyTorch, bf16 (about 24 GB of GPU memory), with text, images and audio. For text only, basal (2.2 or later) also serves it through vLLM (--mode vllm; FP8 checkpoint: basal-omni-1.0-max-FP8), llama.cpp and Ollama (basal-omni-1.0-max-GGUF) and MLX on Apple Silicon (--mode mlx-q8, or the 8-bit port basal-omni-1.0-max-MLX); details: OMNI.md. 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.

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.

Tasks (experimental, basal 2.1)

basal 2.1 adds four task types on the same POST /v1/basal API:

• classify: labels with a rubric each.
• verify: statements about the text; each one is supported, contradicted or not stated.
• compress: keeps the sentences, lines or paragraphs that a question needs.
• extract: fills the fields of a JSON schema, copied exactly from the text with character offsets.

The model does not write text. Each task is expanded into the model's own decisions over one shared input, so every answer comes with calibrated probabilities. Experimental: the API may change in a minor release. Details: TASKS.md.

With the server from the quick start:

curl -s localhost:8000/v1/basal -H 'content-type: application/json' -d '{
  "state": "I was charged twice for my card payment yesterday.",
  "questions": {"intent": {"type": "classify", "none": true, "labels": [
    {"name": "card_payment_duplicate", "rubric": "a card payment was taken more than once"},
    {"name": "card_lost", "rubric": "the card is lost or stolen"},
    {"name": "refund_status", "rubric": "asks about a refund in progress"}]}}}'

The answer (answers.intent, rounded to 4 places):

{"type": "classify", "label": "card_payment_duplicate", "top_label": "card_payment_duplicate",
 "scores": {"card_payment_duplicate": 0.9996, "card_lost": 0.0002, "refund_status": 0.0000, "none": 0.0002},
 "confidence": 0.9996, "automated": {"0.01": null, "0.05": null}, "experimental": true}

basal-omni-1.0-max was not trained on these task formats. Measured with basal-serve 2.1 (eager, bf16, one H100):

• automated: this model's calibration file has no task thresholds, so automated is always null for task answers. Fit task_thresholds on your own labelled data (TASKS.md).
• Limits: currency is the weakest invoice field (62%). extract does not yet return arrays of strings or objects (line items), or text fields read from images.

Images and audio

With --media on, a request may carry images and audio 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" | "audio", "data": "<base64>"} (a data: URL works too). Images: any format Pillow reads. Audio: WAV, FLAC, OGG or MP3, any sample rate (mixed to mono, resampled to 16 kHz), at most 30 s a clip.
• "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.
• Mixed media: one request may mix images and audio, and every question sees all of them. In our test (20 pairs of a document image and a spoken request), adding the other medium changed 10–25% of the answers. For the steadiest answers, send each question with only the media 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 media and from its text):

• Rendered documents are read almost as well as their text (6 points behind), and photos add 13 points over the question alone.
• Speech scores 22–42 points below its transcript and stays at the untrained base's level: send a transcript when you have one.

Speed

One question per request, one option order, on one H100 (--mode eager, bf16, --media on):

• basal-1.5-max in its fast engine: 27.3 ms p50 and 33.4 decisions/s on the same test.
• On Werdykt's 5,000 items (longer states, 16 concurrent clients): 4.2 decisions/s, 76 ms p50 / 1.36 s p95 (basal-1.5-max in its fast engine: 4.6 decisions/s).
• --orders 2 costs about twice as much per question.
• Long states (up to ~30k tokens measured) are prefilled in 2,048-token chunks.

Files

• model-*.safetensors, tokenizer, chat_template.jinja, processor_config.json: the model.
  • Gemma 4 architecture, 12B, 262k vocabulary.
  • Gemma's image and audio encoders are included and unchanged (this release is trained on text); the engine reads media through them with --media on (processor_config.json).
• 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 47.4% / 82.9% of decisions at 0.47% / 3.9% 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-4-12B-it with LoRA (rank 32) on the language-model layers. The vision and audio parts were not trained.
• Data: about 815,000 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: one epoch of 815,386 decisions (12,741 steps of 64; the basal-1.7 mix without its 1,931 items longer than 6,144 tokens) at learning rate 3e-5.

Limitations

• Experimental. It is the first model of a new base; validate it on your own documents before relying on it.
• Images and audio are zero-shot. Gemma 4's encoders read them, but this model was trained on text decisions only; media answers are weaker than text where the media itself is hard (phone photos, speech), see "Images and audio". Their calibration is the text calibration.
• Speed: in the eager engine (needed for images and audio) it is slower than basal-1.5-max in the fast mode; text also runs on vLLM, llama.cpp and MLX (see Quick start).
• 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: max usually answers "no suitable tool", but is often unsure between that and "ask the user" (2 of 6 out-of-scope requests 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-4-12B-it (Apache-2.0).