AIEval DoW Bench v1.1.1 · Last verified Oct 10, 2026
What is not ranked, and why
Oversized: marketed smaller, but over 6B total weights
These models were excluded because, although marketed at a smaller effective size, their total weights exceed the 6B category. Stage 1 counts total weights: every weight has to be loaded, so a smaller effective or active count does not bring a model back in.
| Model | Developer | Total weights | Marketed figure | Why it is out |
|---|---|---|---|---|
| Gemma 4 E4B | Google DeepMind | 7.996B | 4.5B effective | Total weights are 7.996B once per-layer embeddings are counted. |
| Arcee Trinity Nano | Arcee AI | 6.120B | 1.0B active | Mixture-of-experts. All experts load, so the total is the size that counts. Only a preview checkpoint is published. |
License: terms prohibit military use
- Llama 3.2 (1B, 3B). The Llama 3.2 Acceptable Use Policy prohibits military and warfare use.
- Gemma 3 (1B, 4B). Covered by the Gemma Terms of Use and the Gemma Prohibited Use Policy. That policy does not apply to Gemma 4, which is plain Apache 2.0 and is ranked.
Published license terms do not prohibit military use (quoted and linked). Not legal advice; confirm with your contracting officer or SJA.
Origin
The developer must be US-based. Models from non-US developers are not eligible, regardless of license or how easy they are to run locally.
Larger models
Models over 6B total weights are out of scope for v1.1.1, which targets what a unit can run on a single 12 GB workstation GPU.