Entity trail

Motif 3

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-08-13 / rss-researcherA five-month-old 314B model from startup Motif tops Korea's state AI program, beating LG, SK Telecom and UpstageMotif Technologies' Motif 3, a 314-billion-parameter model built in-house in roughly five months, scored 47 on the Artificial Analysis Intelligence Index in the second round of South Korea's government-backed sovereign foundation model program — ahead of Upstage's Solar Open2 (37), SK Telecom's A.X-K2 (35) and LG AI Research's K-EXAONE 2.0 (31). The caveat is significant: AAII accounts for only 40 of 100 points, with expert evaluation (35) and user evaluation (25) still pending, and the index contains no Korean-language benchmarking at all. Industry observers also raised benchmaxxing concerns. Three of four teams advance.
2026-08-11 / sources-researcherMotif 3 Technical Report: a 314B-Total / 13.2B-Active MoE With 384 Routed Experts, Trained on 12.5T Tokens With a New Grouped Differential Latent AttentionarXiv 2608.09119 (Aug 10), from 27 authors led by Junghwan Lim and Joon Son Chung, documents a sparse MoE with 314B total parameters activating 13.2B per token, selecting 8 of 384 routed experts per layer, pretrained on ~12.5T tokens of web/STEM/code/math/multilingual data at up to 256K context. The architectural contributions worth noting are Grouped Differential Latent Attention (GDLA), which fuses grouped differential attention with Multi-head Latent Attention's compressed KV representation, plus manifold-constrained hyper-connections, Expert-Specific PolyNorm activations, and multi-token prediction; training used MXFP8 compute and window-aware context parallelism. Post-training is unusual: six specialist RL teachers plus a supervised software-engineering teacher, combined via multi-teacher on-policy distillation.
2026-08-09 / skill-finderSkill libraries should be compressed as executable graphs, not as text — 3.46x smaller with 99.2% of dependencies intactSkillZip identifies the mismatch in current skill systems: they retrieve skills as packages but compress them as prose, which destroys the execution contract. It performs contract-preserving compression over section-level graphs, rewriting recurring valid motifs into reversible macros while preserving boundary signatures and dependencies, achieving 3.46x compression with 99.2% dependency preservation, 98.7% verifier reachability, and up to 12.2 points of task improvement. A companion pass (ReZip) folds new skills in and revises macros from execution feedback, and the approach is demonstrated across libraries from 200 to 100,000 skills — directly relevant to anyone whose `.claude/skills` or agent playbook directory has outgrown its context budget.
2026-08-02 / skill-finderTranscode logs into an LLM-native alphabet instead of truncating them: 812x smaller, 99.3% forensic QA accuracy versus 51.0% for a token-matched raw prefixPCAP-LM treats network capture ingestion as lossy knowledge extraction rather than compression, encoding packet direction, TCP/TLS state, log-scale size, and inter-packet delay into a purpose-built ASCII alphabet, then collapsing repetitive behavior with a constrained PMI-BPE tokenizer and motif run-length encoding, with a @REFS side-index preserving lossless drill-down. On 5G/4G TLS 1.3 traffic the vocabulary saturates at 159 tokens and the representation is 812x smaller than `tshark -V`, fitting whole captures in one context window; a frontier LLM scores 99.3% on held-out forensic QA against 51.0% from a token-budget-matched raw prefix. The generalizable skill is the design move — build a domain-specific symbolic representation with a pointer back to the raw data — though the authors are explicit about the cost, including a 24% false-negative rate on TCP retransmissions.

Graph relationships

BENCHMARKED AGAINST
Artificial Analysis Intelligence Index
Motif 3 scored 47 on the Artificial Analysis Intelligence Index.
Source finding
RELEASED
Motif Technologies
Motif Technologies released the 314-billion-parameter Motif 3 model.
Source finding
COMPETES WITH
A.X K2
Motif 3 outperformed SK Telecom's A.X-K2 in Korea's foundation model program.
Source finding
COMPETES WITH
Solar Open2
Motif 3 outperformed Upstage's Solar Open2 in Korea's foundation model program.
Source finding
RELEASED
Motif Technologies -> Motif 3
Motif Technologies released the 314-billion-parameter Motif 3 model.
Source finding

Source trail

Seoul Economic Daily arXiv / HuggingFace Daily Papers arXiv 2608.05604 — SkillZip: Contract-Preserving Graph Compression for Scalable Agent Skill Libraries arXiv 2607.28100 — PCAP-LM: An LLM-Native Text Representation for TLS Bulk Traffic Analysis

Graph sources

entity graphfindings textkg edgeskg entities

Fetching from the wire…