Nagi-ENORMOUS · 27B

Nagi-ENORMOUS is the fourth model line in Nagi, after SMOL (0.5B), BIG (4B) and HUGE (12B). It scores a supplied closed set of options in one forward pass and returns a choice plus option probabilities. It does not generate a chain of thought or a free-form answer.

This repository distributes the rank-8 LoRA adapter for the public Apache-2.0 Qwen3.8-27B base, pinned at 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. It is not a standalone merged weight bundle; the SDK downloads the base separately.

Nagi-ENORMOUS is the strongest Nagi line. It placed first in Arena Live, our real-time game benchmark: 83/90 points and 23/30 rounds won against Jev 1.13.0, OpenJev and Laya. It is statistically tied with Jev1.13.0 on the public four-system decision suite.

Public four-system decision suite

Same protocol and scoring as the Nagi-HUGE release. Typed-choice adaptations of 8 public sources, 4,671 core examples, each source weighted equally.

• ENORMOUS − Jev: −1.20 pp [−2.70, +0.31]. The interval includes zero, so the two are statistically tied on this suite.
• ENORMOUS − HUGE: +5.27 pp [+3.49, +7.14].
• By source, ENORMOUS is ahead of Jev on XNLI in all four languages (en 93.5, vi 83.0, de 86.0, zh 79.5), ANLI, BoolQ, CB and RTE, and behind on MMLU-Pro (65.0 vs 82.3).
• The option-order flip rate was not measured for ENORMOUS.
• This is not an official leaderboard score, and exposure to these sources during pretraining is unknown.

Arena Live: real-time games

We ran three games with real rules and physics. Every model got the same observations and options, moves were simultaneous, seeds were pre-committed, seats rotated and play was lockstep. Each game had 10 rounds, scored 3/2/1/0 points per round (max 90 points, 30 round wins).

• Rotorwash (2D rigid-body helicopter): ENORMOUS and Jev tie on points, but ENORMOUS stayed airborne longer in 9/10 rounds (median 35.8 s vs 13.8 s).
• Lightcycle Royale: ENORMOUS won 8 of 10 rounds.
• Stack Attack: ENORMOUS won 10 of 10 rounds.
• Among Nagi lines: 7 → 39 → 65 → 76 points for SMOL → BIG → HUGE → ENORMOUS.
• Scope: exhibition scale (10 rounds per game); every model got the same shared sensor features. Arena Live measures skill under fixed, known rules. On our internal test of reading brand-new rules, the 27B line does not yet improve over the 12B line.

Videos: https://nagisanzenin.github.io/nagi/arena-live/

Use

python -m pip install "nagi-decisions @ git+https://github.com/nagisanzenin/nagi.git"

from nagi import load_enormous
model = load_enormous(device="cuda")
result = model.system_one(
    state={"A": 7, "B": 3},
    questions={"decision": {"type": "choice", "instructions": "Which value is greater?",
                            "criteria": {"A": "A is greater", "B": "B is greater"}}},
)
print(result["answers"]["decision"])

• Validated path: unmerged LoRA, CUDA BF16, eager attention, the official Qwen chat template with enable_thinking=False, and selected option-letter projection.
• Temperature: 1.0, which is not a fitted calibration.
• Options and length: choice supports 2–26 options. Training prompts were ≤768 tokens; longer inputs are outside the trained range. Rendered input is capped at 4,096 tokens and fails with an explicit error, with no silent truncation.
• Memory: the BF16 weights need about 56 GB, so plan for an 80 GB GPU.

Training and selection

• Recipe: HUGE's. A 75-update pilot, then continued supervised LoRA (effective batch 8, AdamW learning rate 3e−5), rank 8, alpha 16, no dropout.

• Selection: release rungs at 125, 250 and 500. The preregistered stop rule ended training at step 500 (dev policy 0.9397, language 0.8631), and step 500 was selected on dev only.

• Evaluation data: no public evaluation gold and no final-set data were used for selection.

• Adapter weights SHA256: b8d307527a0426fd8208bf3c44eebaf47f14057452403ee282c08fe72053ddf7

• Adapter config SHA256: 9b9529fcf29318ff348abfd2db9face3ecdd0b9bf281734cec739ec186d86c09

Links

SDK and documentation · Nagi-HUGE · Nagi-BIG · Nagi-SMOL