Motif 3
modelYour notes
Motif Technologies' frontier Korean/English foundation model: a 314B-total / 13.2B-active sparse MoE (384 routed experts, top-8, plus 1 shared expert; 53 layers, 2 dense + 51 MoE; 220K vocabulary) with a native 256K context, pretrained from scratch on ~12.5T tokens (knowledge cutoff March 2026; ~70% of the corpus is NVIDIA's public Nemotron pretraining releases, the rest in-house web, STEM, code, synthetic, Korean, legal and financial data). The architecture is fully in-house: Grouped Differential Latent Attention (GDLA) (grouped differential attention over an MLA-style compressed KV latent, with a query-dependent output gate), Expert-Specific PolyNorm activations, modified manifold-constrained hyper-connections (mHC, 4 residual streams), and a 1-layer multi-token-prediction head used for self-speculative decoding. Trained on NVIDIA B200s with the Muon optimizer (QK-Clip), selective MXFP8 expert compute and window-aware context parallelism. Post-training combines general SFT, six RL-trained specialist teachers plus a software-engineering teacher, and Multi-teacher On-Policy Distillation (MOPD) into one unified reasoning model.
Timeline. Motif first surfaced as Motif-3-Beta, an intermediate checkpoint released 14 July 2026 under a non-commercial research-only license (AA lists it separately as "Motif 3 (Beta)"; it scored 44 then 45 on AA Intelligence Index v4.1/v4.1.1, 37 on v4.2, and 32 on v4.3). The final release followed in August 2026: the post-trained Motif-3 and pretrained Motif-3-Base checkpoints under the MIT license (commercial use permitted, no access gate), the "Motif 3: Technical Report" (arXiv 2608.09119, 10 Aug), and an NVFP4-quantized build for Blackwell GPUs. AA lists the final model at AA Intelligence Index v4.3: 34 — the #1 Korean model on the index — with particular strength on agentic tool use (τ³-Banking 35.3, Terminal-Bench 2.1 74.9, ITBench 51.5) and calibrated abstention on AA-Omniscience (71.6 non-hallucination). Motif's torchtitan-based training framework (motif3-training-example) and a vLLM fork with Motif-3 support are open on GitHub. From a lab spun out of Moreh, trained on its GPU-optimization stack.
Model Details
Benchmark Scores
| Benchmark | Score | Mode |
|---|---|---|
| GDPval-AA v2 | 38.7 | — |
| τ²-Bench Telecom | 94.7 | — |
| τ³-Banking | 35.3 | — |
| ITBench | 51.5 | public subset |
| SWE-bench Verified | 76.2 | — |
| Terminal-Bench 2.1 | 74.9 | — |
| SciCode | 40.6 | — |
| IMOAnswerBench | 83.2 | — |
| Apex-Shortlist | 75.5 | — |
| GPQA Diamond | 83.4 | — |
| Humanity's Last Exam | 37.0 | — |
| CritPt | 6.6 | — |
| AA-Omniscience | 30.1 | accuracy; 71.6 non-hallucination |
| AA-LCR | 72.3 | — |
| IFBench | 78.2 | — |
| MMLU | 86.2 | Motif-3-Base, 5-shot |
| MMLU-Pro | 68.6 | Motif-3-Base, 5-shot CoT |
| GSM8K | 93.9 | Motif-3-Base, 8-shot CoT |
| HumanEval | 73.7 | Motif-3-Base, 0-shot |
Variants
| Name | Parameters | Notes |
|---|---|---|
| Motif-3 | 314B | Final post-trained release (Aug 2026), MIT license, ships the MTP head for self-speculative decoding |
| Motif-3-Base | 314B | Pretrained base checkpoint (no SFT/RL, no chat template, MTP head not shipped), MIT license |
| Motif-3-NVFP4 | 314B | NVFP4 weight-quantized build of Motif-3 for NVIDIA Blackwell; within ~0.15 pts of the Block-FP8 release on Motif's suite |
| Motif-3-Beta | 314B | July 2026 intermediate checkpoint, non-commercial research-only license; AA slug motif-0714 ("Motif 3 (Beta)"), AAII v4.3 32 |