Saina Helm 3

Helm 3, natively multimodal. Saina Helm 3 reads text, images and audio together in one pass and answers typed questions about them: yes/no, single choice, multiple choice or rating. It returns a probability for every answer instead of generating text, so every decision comes with a confidence you can act on.

What it does

• One pass, many questions. Send the state, your images and audio clips, and any number of questions. Every question is answered from a single pass over the shared input.
• Reads everything at once. Text, screenshots, photos, charts, receipts and short audio clips arrive in the same request, and every question sees all of it.
• Knows when to abstain. In decision mode, low-confidence answers come back empty with a reason, so your system can escalate instead of guessing.
• Small and fast. 0.8B parameters; runs on a single GPU, on Apple silicon, or on CPU.

Benchmarks

Saina Helm 3 scores 45.89 on the Decision Index, a benchmark of decision tasks (public index, full 140,620-request suite). Scores are out of 100: each benchmark is chance-corrected, so 0 means random guessing and 100 means perfect.

Full runs, self-run with the published harness (run-saina/saina-decision-index).

Intended use

Decisions where a wrong answer is recoverable or reviewed before it takes effect: routing a support message with its screenshot, checking a receipt against a claim, flagging a call whose voice may be synthetic, tagging sounds and images. Validate on your own labelled data, for your own label space, before use.

Out of scope

• Generating text of any kind; the model only scores the options you supply.
• Final, unreviewed decisions with material consequences for a person (credit, employment, medical, legal).
• Identifying or classifying people, or profiling individuals.
• Safety filtering or content moderation.
• Video, and audio longer than 10 seconds.
• Decisions taken directly on untrusted content: text, images and audio can carry instructions that steer the answers.

Usage

pip install "saina[local,media]>=0.9.1"

from saina.transformers import HelmModel

helm = HelmModel.from_pretrained('run-saina/saina-helm-3-0.8b', device='cuda:0')

answers = helm.ask(
    model='saina-helm-3-0.8b',
    state={'claimed_usd': 244.80, 'policy': 'Client meals up to $300 with an itemized receipt.'},
    images=['receipt.png'],
    questions={
        'total_matches': {'type': 'yes_no', 'question': 'Does the receipt total match the claimed amount?'},
        'action': {'type': 'single_choice', 'question': 'How should this claim be handled?',
                   'options': {'approve': 'Approve', 'review': 'Send to a reviewer', 'reject': 'Reject'}},
    },
)

Use the saina package: it builds the exact prompt the model expects and applies its answer head. This model needs saina 0.9 or newer (older versions refuse it); use 0.9.1 or newer, which uses less memory on long inputs.

Where to run it

Inputs

Question types

Pass mode='decision' with a threshold to get abstentions instead of low-confidence answers.

License

Apache 2.0. Built on Qwen/Qwen3.5-0.8B.