Suspected Frontier Model Architecture: No RoPE and Hybrid Attention
nrehiew_ · x · 2026-07-16
The post analyzes the architectural details of a suspected frontier model. Regarding the attention mechanism: it adopts a sliding window and uses 1D convolution on KV and residual connections; RoPE is completely abandoned in favor of distance-based relative position bias. On the MoE side: total parameters are 975B with 41B active, trained on 45T multimodal tokens; it uses an architecture similar to DeepSeek v3 but features 6 routed experts and 2 shared experts (usually 1 or 0), thereby increasing the sparsity rate.
Related event: Inkling Performance and Related Architecture Speculation(7 posts)→
More from Research
- Open-source decision models sold their own base in Red Alert 2 — then learned until no human could beat them — heathercmiller · 2026-10-27
- DeepMind paper: LLMs can't sample randomly on their own — pass random numbers from code instead — rohanpaul_ai · 2026-10-27
- World Model Assigns Weight and Friction to Objects, Then Simulates with a Physics Engine — jonstephens85 · 2026-10-27
- Token Collapse: Study Finds Graph Soft Prompting Tokens in LLMs Fail to Encode Structure — Sapienza · 2026-10-27
- Mathathon 2.0 launches with Terence Tao's SAIR, $1.5M in compute for teams — _sathvikr · 2026-10-27
- The 'sweet lesson' is real: two kinds of ML problems need opposite strategies — mike64_t · 2026-10-27