AIAny
Icon for item

Motif 3: Technical Report

Describes a 314B-parameter decoder-only Mixture-of-Experts language model that activates 13.2B parameters per token for fine-grained sparsity, long-context (up to 256K) and multi-domain capabilities. Emphasizes GDLA architecture, expert balancing, and multi-teacher distillation.

Introduction

Most scaling work trades compute for dense parameter counts; Motif 3 pursues a different point on that tradeoff by packing far larger expert capacity into each layer while keeping per-token compute bounded. That lets the model target very long contexts and specialized capabilities (reasoning, coding, scientific knowledge) without an equivalent increase in FLOPs per token.

Key Findings
  • Large sparse capacity with limited compute: each MoE layer has 384 routed experts with 8 experts selected per token, yielding 13.2B activated parameters while keeping computation manageable — so what: this provides much higher effective capacity for specialization without proportional per-token compute growth.

  • Architecture innovations for stability and specialization: introduces Grouped Differential Latent Attention (GDLA), Expert Specific PolyNorm activations, and modified hyper-connections — so what: these changes aim to improve expert specialization and numeric stability during large-scale sparse training.

  • Training and inference scale: pretrained on ~12.5T tokens across web, STEM, code, math, and multilingual corpora, with infrastructure choices (selective MXFP8, fused kernels, context parallelism) enabling context lengths up to 256K — so what: supports long-horizon tasks and agentic workflows that require persistent memory over very long documents.

  • Post-training consolidation: combines supervised fine-tuning, six specialist RL teachers, a software-engineering teacher, and Multi-teacher On-Policy Distillation — so what: unifies complementary capabilities (reasoning, coding, tool use) into a single model without separate specialist deployments.

Who it's for and tradeoffs

Great fit if you need a single-model solution that handles long-context understanding and a mix of reasoning/coding/scientific tasks and can accept the engineering complexity of sparse MoE (routing, expert balancing, specialized kernels). Look elsewhere if you require minimal engineering overhead, strict determinism for inference on commodity hardware, or only short-context dense LLM behavior.

Where it fits

Positioned among large open-weight foundation models that explore sparsity as a path to capacity — similar aims as other MoE and long-context efforts, but distinguishes itself by: combining very fine-grained expert routing, novel attention variants (GDLA), and an extensive multi-teacher distillation pipeline to consolidate capabilities.

Method notes

Key implementation details include 314B total parameters, 13.2B activated per token, grouped differential latent attention with compressed key-value representations, expert-balancing and numeric-stabilization strategies for stable training, and system-level optimizations (MXFP8 selective compute, memory-efficient fused kernels) to support extremely long contexts and large-scale sparse training.

Information

  • Websitearxiv.org
  • AuthorsJunghwan Lim, Joon Son Chung, Sungmin Lee, Wai Ting Cheung, Gihun Cho, Minsu Ha, Sangho Kang, Beomgyu Kim, Dongseok Kim, Jangwoong Kim
  • Published date2026/08/10

More Items

Transfers hierarchical, structured memory from a large teacher agent into small LLM agents to improve tool-use success. Constructs Workflow, Subtask and Function memories with proactive/reactive injection; training-free and validated on multiple tool-use benchmarks.

A research report proposing a continual-learning agent workflow that pairs recursive self-improvement with a Mixture-of-LoRA design: freeze a foundation model, compose specialist LoRA adapters routed per user turn, and support them with long-context RL and post-training infrastructure.

Analyzes why supervised fine-tuning (SFT) causes severe task conflicts under multi-stage multi-task training while reinforcement learning (RL) enables stable coexistence, attributing the effect to sparse, near-orthogonal RL parameter updates and proposing Parallel-RL to decouple multi-task training.