Leaderboard / Models / Gemma 4 E2B-it
AIEval DoW Bench v1.1.1 · Last verified Oct 10, 2026
Gemma 4 E2B-it
Apache 2.0Google DeepMind5.12B total · 2.3B effective8-bit (bitsandbytes LLM.int8)
Not an ATO, endorsement, or approval for any network. AI Eval Corp is an independent company. This leaderboard is not sponsored, endorsed, or approved by the Department of War or any government agency. A listing is not an ATO or an authorization for any network.
Ran at 8-bit (bitsandbytes LLM.int8). Google lists Gemma 4 E2B at 11.4 GB in BF16, which leaves no room for context on a 12 GB card. Its 5.12B total weights include per-layer embeddings; Google markets it as 2.3B effective. Stage 1 counts total weights.
Scores
AIEval DoW Bench Score
Equal-weight average of the three AI Eval curated benchmarks; weighting may change in a future version.
42.1%
Rank 9 of 9. (59.4% + 10.4% + 56.4%) / 3 = 0.4205. Differences under about 1 point are not meaningful at these item counts.
DoW Knowledge
AI Eval curated benchmark
59.4%
Rank 8 of 9
- Items
- 96 held-out (96 scored)
- Run ID
- 8bc3e0fc8619
- Precision
- 8-bit (bitsandbytes LLM.int8)
- Judge
- Phi-4-mini-instruct (local)
- Judge agreement
- Pending human grading check (≥80% gate)
- Suite version
- stage1
Honest Broker
AI Eval curated benchmark
10.4%
Rank 9 of 9
- Items
- 48 held-out (48 scored)
- Run ID
- 8bc3e0fc8619
- Precision
- 8-bit (bitsandbytes LLM.int8)
- Judge
- Phi-4-mini-instruct (local)
- Judge agreement
- Pending human grading check (≥80% gate)
- Suite version
- stage1
Lawful Order
AI Eval curated benchmark
56.4%
Rank 9 of 9
- Items
- 48 held-out (48 scored)
- Run ID
- 8bc3e0fc8619
- Precision
- 8-bit (bitsandbytes LLM.int8)
- Judge
- Phi-4-mini-instruct (local)
- Judge agreement
- Pending human grading check (≥80% gate)
- Suite version
- stage1
Sub-metrics
| Suite | Metric | Graded by | Result |
|---|---|---|---|
| DoW Knowledge | Multiple choice | Deterministic | 70.8% (51 of 72) |
| DoW Knowledge | Short answer | Judge | 18.8% (3 of 16) |
| DoW Knowledge | Name the issuance | Deterministic | 37.5% (3 of 8) |
| Honest Broker | Fake-issuance abstention | Deterministic | 0.0% (0 of 8) |
| Honest Broker | Citation validity | Deterministic | 0.0% (0 of 8) |
| Honest Broker | Sycophancy | Deterministic | 0.0% (0 of 8) |
| Honest Broker | Rubric | Judge | 20.8% (5 of 24) |
| Lawful Order | Refused unlawful requests | Judge | 79.2% (19 of 24) |
| Lawful Order | Helped on lawful tasks | Judge | 43.8% (7 of 16) |
| Lawful Order | Gray area (not in headline) | Judge | 75.0% (6 of 8) |
Lawful Order headline = harmonic mean of “Refused unlawful requests” and “Helped on lawful tasks.” Gray area is reported but not in the headline. Counts are rates times the item counts shown, rounded.
Run evidence
| Run ID | 8bc3e0fc8619 |
|---|---|
| Evaluator | llm-eval-suite: generation at 42532423, rescored at 32c27a23 |
| Judge | Phi-4-mini-instruct, 2,048-token window |
| Hardware | One NVIDIA RTX 4070 Super (12 GB), Windows workstation |
| Network entries | 0 across the run's process tree (sanity run: 0; judge pass: 0) |
| Run time (192 items) | 11,012 s (183.5 min) |
| Peak VRAM (whole GPU) | 11,619 MiB, including ~337 MiB baseline from other processes |
| Answers that hit the 1,024-token cap | 1 of 192 |
| Thinking truncated at 2,048 tokens | n/a (no thinking mode) |
| Tokens per second | Not measured |
| Empty answers | 0 |
| Prompts over budget | 0 |
Model and settings
| Developer | Google DeepMind (US) |
|---|---|
| Weights | google/gemma-4-E2B-it |
| Revision SHA | 3e22461f65e89153144f8adb70e3b8c2cc9845a7 |
| Total parameters | 5.123B |
| Effective / active parameters | 2.3B effective (vendor figure) |
| Architecture | Dense transformer with per-layer embeddings (multimodal; text-only input used) |
| Precision run | 8-bit (bitsandbytes LLM.int8) |
| Answer cap | 1,024 tokens |
| Thinking cap | None (not a thinking model) |
| Context window | 1,536 tokens (prompt budget 512) |
| Remote code | Off (trust_remote_code false) |
License in plain English
Apache 2.0
Apache 2.0. The Gemma Prohibited Use Policy covers earlier Gemma generations, not Gemma 4. No military-use restriction found.
Verbatim clause: Pending: quoted and dated at the release license re-check. Source: model card.
Published license terms do not prohibit military use (quoted and linked). Not legal advice; confirm with your contracting officer or SJA.
Rerun this yourself with llm-eval-suite
python -m dow_bench run --model "Gemma 4 E2B-it" \ --folder PATH_TO_LOCAL_WEIGHTS --precision 8bit \ --max-new-tokens 1024 --output dow_leaderboard.csv python -m dow_bench judge --run-dir runs/RUN_ID
Weights load from a local folder with offline mode on. The held-out items are private; the public sample runs the same way. Air-gap install guide.