LEV (OpenLAIR)
Quadrant scores
See the full quadrantScored with the public quadrant rubric (maturity × capability; bubble = adoption). Revise as evidence lands.
- Maturity4.3/10
Apache-2.0 LoRA + pointer head on Ouro-1.4B, Ouro-2.6B and Nanbeige4.2-3B, with cards, calibration files, GitHub examples, datasets and a paper. Card says research use only. No server, package, integrations or SLA.
- Capability4.4/10
Probabilities over caller-provided options only (choice). Per-depth temperatures, no ECE in the card. Lev-2.6B 63.2% → 76.2% on JevBench's 231 public tasks at depth 1 → 4 (authors). p50 51–200 ms on one H200 (authors).
- Adoption1/100
About 200 downloads and 4 likes across the three HF repos (25 Oct); 105 upvotes on HF Papers; HuggingPapers post 24 Oct. No third-party integration or production use found.
Vendor claims
- Lev-2.6B goes from 63.2% at depth one to 76.2% at depth four on JevBench (231 public tasks, fp32 readouts)[Vendor claim — not independently verified]
- On-demand depth saves 15% to 24% of the computation at a cost of under one accuracy point[Vendor claim — not independently verified]
- H200 bf16 batch-one p50 51.1 / 100.7 / 150.4 / 200.1 ms at depths 1–4 (Lev-2.6B)[Vendor claim — not independently verified]
publicado (OK do Fabricio 25/10 BRT). Scores are calculated from scratch.
Not the same project as Lev (Interfaze, a LoRA on Qwen3.5-4B). Same name, different authors and method.
What it is
LEV (paper, arXiv 2610.20951, 14 Oct 2026; GitHub; models) turns a pretrained looped language model into a typed decision model. The backbone reruns one stack of layers several times. LEV reads the options after every loop with a shared pointer head and trains all depths together, so one checkpoint answers at depth 1, 2, 3 or 4. More depth means more compute but no generated tokens.
| Model | Backbone | Depths | JevBench public, first → deepest (authors) |
|---|---|---|---|
| Lev-1.4B | Ouro-1.4B | 1–4 | 58.0% → 68.0% |
| Lev-2.6B | Ouro-2.6B | 1–4 | 63.2% → 76.2% |
| Lev-3B | Nanbeige4.2-3B-Base | 1–2 | 69.3% → 72.7% |
Each repo holds the seed-0 LoRA adapter, decision head, tokenizer and per-depth temperatures. The code is Apache-2.0 and derives from Kev. Same idea as SanSi (looped readout on Ouro), from a different group.
Limits (read before you ship)
- Options only. You pass a question and its options and get a probability per option. There is no separate yes/no or score type and no
/v1/systemoneserver. - The authors say gains depend on the workload: Lev-2.6B failed their locked LevBench G3 criterion in all three seeds, and a deeper pass can overturn a correct shallow answer.
- The upstream card says “research purposes only”. Timing is H200, bf16, batch one; accuracy uses fp32 readouts.
Author claims (UnverifiedClaim)
- Lev-2.6B: 63.2% → 76.2% on JevBench’s 231 public tasks, depth 1 → 4 (source)
- On-demand depth saves 15–24% of compute for under one accuracy point (source)
- p50 51–200 ms at depths 1–4 on one H200 (source)
No row on the Decision Index, the JevBench board or Artificial Analysis.
Scores (from scratch, 25 Oct 2026)
- Capability 4.4 = types 3 × 0.30 + calibration 4 × 0.25 + latency 6 × 0.25 + accuracy 5 × 0.20 = 0.90 + 1.00 + 1.50 + 1.00 = 4.40.
- Maturity 4.3 = availability 7 × 0.30 + docs 6 × 0.25 + integrations 1 × 0.25 + support 2 × 0.20 = 4.25 → 4.3.
- Adoption 1 = engagement 2 × 0.35 + ecosystem 0 × 0.35 + production 0 × 0.30 = 0.70 → 1.