Nano Preview: 6.120B total / ~1B active; Mini: 26.124B total / ~3B active (MoE)
131,072 tokens (128K)
Apache 2.0 at release; OpenMDW-1.1 since May 28, 2026
Two text-only, decoder-only AFMoE models trained from scratch in the U.S.: experimental, personality-forward Nano Preview for low-resource chat and instruction following, and fully post-trained Mini for reasoning, tool calling and multi-step agent workflows. Nano has 56 layers and Mini 32; both use 128 routed experts with 8 selected per token plus one shared expert, two initial dense layers, gated grouped-query attention with QK normalization, and a 3:1 pattern of 2,048-token local RoPE attention and global NoPE attention. Each was pretrained on 10T Datology-curated tokens in 7T/1.8T/1.2T phases with increasing math and code concentration, using Muon for hidden layers, AdamW for embeddings/output, modified TorchTitan and HSDP on up to 512 H200 GPUs. Nano underwent 256K-sequence context extension for 128K inference; Mini was extended at 128K. The released Nano Preview is not a thinking model and may be unstable in edge cases. Mini is reasoning-tuned, but the released Nano/Mini post-training objectives and budgets are not fully disclosed; the later family report's SFT/RL section concerns Trinity Large, not these checkpoints. Arcee reports Mini scores of 84.95 on zero-shot MMLU, 92.10 on Math-500, 58.55 on GPQA-Diamond and 59.67 on BFCL V3, alongside 63.49 on MUSR and 8.90 on SimpleQA; sampling, tool scaffolds, retry budgets and aggregation are unspecified. Base, pre-anneal and later Large/precision variants are not silently grouped with this two-model launch.