Lev1 (K2-Horizon-3.7B)
Lev1 is a System One decision model, in the line of Typesafe AI's Jev: fast, intuitive calls on structured input. You send a state and typed questions, and Lev1 returns a probability for every option in one forward pass. It never generates text.
- Fast. 28.8 ms median, 169.9 ms p95 per request on one RTX PRO 6000 Blackwell Server Edition.
- Scored. 42.28 on Decision Index 0.2.1, from a complete run with the official harness.
- Best in class on tools. 68.8 on Tools & Automation, above every model under 10B parameters on the board (as of 2026-09-28).
- Small and open. A 3.7B model: a LoRA on IFM/K2-Horizon-3.7B, shipped as merged weights and as the adapter, under Apache-2.0.
- Clean data. Training sources carry commercial-use licences and passed decontamination against the benchmark suite. The synthetic part is public as theaviv/lev1-decisions.
Training
I trained a LoRA (rank 32, alpha 64) on questions sampled from a 163,611-question mixture. Each short question also appears in a second option order, and a consistency loss pulls the two answers together. For 15,591 questions the target blends the gold label with Gemma-4-31B-it's probabilities, on sources where Gemma matched gold at least 60% of the time. MiMo-V2.6-Flash wrote the 4 LLM-written sets; I kept a text only when Qwen3-235B-A22B agreed with its label, and Gemma-4-31B-it too for all but the phishing emails.
| Data | Licence | Role | Train questions |
|---|---|---|---|
| theaviv/lev1-decisions (17 generators, 4 LLM-written sets) | CC-BY-4.0; CC-BY-SA-3.0 (hallucination set) | synthetic | 31,837 |
| Open-Jev | CC0-1.0 | general | 43,311 |
| multi_nli, snli, boolq, contract-nli, SemEval_NLI4CT, RAGTruth-processed | OANC/CC-BY/CC-BY-SA/MIT; CC-BY-SA-4.0; CC-BY-SA-3.0; CC-BY-4.0; MIT (annotations); Summary task only (news contexts) | inference | 19,686 |
| winogrande, social_i_qa, piqa, openbookqa, ai2_arc | Apache-2.0; CC-BY-4.0; AFL-3.0; CC-BY-SA-4.0 | commonsense | 18,315 |
| esci, HelpSteer2, newyorker_caption_contest, chess-puzzles | Apache-2.0; CC-BY-4.0; CC0-1.0 | mixed | 16,950 |
| clinc_oos, banking77, hwu_64, amazon_massive_intent | CC-BY-3.0; CC-BY-4.0 | intent detection | 13,061 |
| esci | Apache-2.0 | retrieval | 12,280 |
| gsm8k | MIT | reasoning | 6,026 |
| When2Call | CC-BY-4.0 | tool use | 2,145 |
Results
| Lev1 | Decision 1.0 Lux | JPT-4B | Jet v6.2 | Decider 4B | Kev 4B | |
|---|---|---|---|---|---|---|
| Decision Index | 42.28 | 43.49 | 43.04 | 42.60 | 40.70 | 34.64 |
| Knowledge & Reasoning | 22.7 | 30.9 | 28.7 | 28.7 | 25.7 | 22.9 |
| Language Understanding | 44.0 | 48.0 | 52.5 | 43.9 | 46.0 | 35.3 |
| Retrieval & Classification | 47.1 | 50.0 | 45.0 | 48.2 | 44.7 | 41.0 |
| Tools & Automation | 68.8 | 57.2 | 57.2 | 62.9 | 58.6 | 52.6 |
| Arts & Human Taste | 30.5 | 26.4 | 25.8 | 27.0 | 25.0 | 17.9 |
The other columns are entrants near Lev1 on the board. Scores are chance-corrected skill: 0 is guessing, 100 is perfect. The run answered all 150,759 requests; the harness scores 150,317 of them and excludes 442.
Try it
import sys
from huggingface_hub import snapshot_download
path = snapshot_download("theaviv/Lev1-3.7B")
sys.path.insert(0, path) # the repo ships its scorer as the `lev1` package
from lev1.engine import Lev1Scorer
scorer = Lev1Scorer.from_pretrained(path) # adapter=True loads the base plus the LoRA instead
probs = scorer.score(
{"channel": "email", "message": "Checkout has returned a 500 error for an hour and orders are failing."},
{
"urgent": {"type": "noul", "instructions": "This message needs a reply within the hour."},
"team": {
"type": "choice",
"instructions": "Which team should handle it?",
"criteria": {"billing": "Payments, refunds and invoices.", "engineering": "Outages and bugs.", "sales": "Pricing and new accounts."},
},
},
)
print(probs["urgent"]["yes"], probs["team"])
Decision Index harness (bring your own copy of the suite):
PYTHONPATH="$(python -c 'from huggingface_hub import snapshot_download as d; print(d("theaviv/Lev1-3.7B"))')" python -m decision_index pipeline --engine lev1.engine:Lev1Engine --edition 0.2.1 --out runs/lev1
Limits
- Two question types,
choiceandnoul, up to 255 options. A request that does not fit raisesUnsupported; Lev1 never truncates or drops an option. - States above 16,384 tokens take a slower chunked path.
- English only. Weakest on Knowledge & Reasoning (22.7) and Arts & Human Taste (30.5).
- The probabilities come straight from the softmax, without a fitted temperature. Check calibration on your data before you set thresholds.
Apache-2.0. Built on IFM/K2-Horizon-3.7B (Apache-2.0, revision fe504ef19c7e). By Aviv Dozorets.
Lev1 is not lev, the Interfaze entrant on the Decision Index board.