Nagi-HUGE · 12B

Nagi-HUGE is the third model tier in Nagi (Smol, Big, Huge), not a rename of the historical Big v3 checkpoint. It scores a supplied closed set of options in one forward pass, returning a choice and option probabilities. It does not generate a chain of thought or free-form answer.

This repository distributes the rank-8 LoRA adapter for the public Apache-2.0 Gemma 4 12B instruction base, pinned at 707f0a3b8a3c7ad586ed01e27eafbad8a27dd0f7. It is not a standalone merged 12B weight bundle. The SDK downloads the base separately.

Research release. No demonstrated general or universal superiority over Jev. The preceding preregistered synthetic release gate failed: policy accuracy Nagi85.83%, Jev88.33%, difference−2.50 percentage points, paired95% interval[−5.83,+0.67]. It is released as a research artifact with the unfavourable evidence preserved. Full internal report and receipts.

Use

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

from nagi import load_huge
model = load_huge(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"])

The validated path is unmerged LoRA, CUDA BF16, eager attention, exact official Gemma chat template with thinking disabled, and selected option-letter projection. Temperature1.0 is not fitted calibration. Choice supports2–26 options; total rendered input is capped at4096 tokens with an explicit error and no silent truncation. The SDK also exposes score/noul through the existing renderer; the new public comparison evaluates choice, not every typed behavior.

CUDA H100 was tested. 12B BF16 weights alone require approximately24GB; activations, loading and allocator overhead need additional memory. No24GB-device compatibility claim or quantized deployment validation. CPU uses FP32 and is not represented by the latency numbers. Base weights are downloaded from Hugging Face on first load.

Training and selection

A75-update pilot preceded the release campaign. This adapter is the dev-selected125-additional-update rung of a continued supervised LoRA phase (effective batch8, serial microbatch1; half policy and half controlled language presentations; AdamW learning rate3e−5). Rank8, alpha16, no dropout. Campaign corpus10,000train/1,000dev; separate frozen synthetic final2,400rows. The250 rung regressed on dev and triggered preregistered early stop. No public evaluation task gold was used to select this checkpoint. Full corpus history and exact configs are in the linked research repository; public pretraining exposure remains unknown.

Adapter weights SHA256: 6ae25683bf309b979dde37995a2371feccaabe64a7e60529e1f0cd44003a4786. Adapter config SHA256: ecc81c85aa8124b893f49b597cb41490dfa5cf8dc333fb0f2ce7c293be5d4fe6.

Speed and scope

A separate warm H100 loopback HTTP probe measured P50/P95 143/145ms at1024tokens and659/664ms at4096tokens,30observations/tier, concurrency1. It excludes WAN/model load and uses length filler; it is not a service SLA, long-context quality test, or matched-hardware comparison with Jev. Jev service hardware is unknown.

Public four-system evaluation

Nagi−Jev: -6.47 percentage points (95% interval [-8.13, -4.89]) on the common-evidence eight-source macro. The interval supports a Jev advantage on this specific suite. It does not establish universal generalization or a same-hardware latency win.

Eight equally weighted public sources; 4,671 core examples and320 option permutations. Primary common evidence uses4,518 upfront-frozen rows; local tokenization is audited, Jev internals unknown. All19,964 model outputs are valid. Nagi developers ran the evaluation. Public pretraining exposure is unknown; these adaptations are not official leaderboard scores.

Separate WAN/API measurement: Jev P50/P95 1086/2962ms, client concurrency4, unknown server hardware. No matched-hardware ranking.

Full protocol, per-source accuracy, paired intervals, calibration, raw receipts and limitations.

SDK · Research · Frozen public protocol