Deep Dive into a Model's Attention and MoE Details
nrehiew_ · x · 2026-07-16
This post breaks down a model's architectural design, focusing on two main areas:
Attention Design
- Uses sliding window, which is somewhat surprising given the team's strong background in linear attention.
- Applies basic 1D convolution on KV and residual.
- No RoPE is used; instead, attention weights are formed by projecting the hidden state and combining it with distance-dependent position bias.
MoE Design
- The model scale is about 975B parameters with 41B active, trained on over 45T multimodal tokens.
- Compared to Dsv3, active parameters increased by roughly 10%, and token count is 3 times larger.
- Features 6 routed experts out of 256 experts, plus 2 shared experts; the author notes that having 2 shared experts is quite rare, as the norm is typically 0 or 1.
Overall, a highly detailed architectural observation post.
Related event: Inkling Performance and Related Architecture Speculation(7 posts)→
More from Models
- BullshitBench update: GPT-6-Astra beats all prior OpenAI models but still trails Anthropic — scaling01 · 2026-09-11
- Astra Scores 83% on GauntletBench, First Computer-Use Agent to Beat Human Baseline — ducha_aiki · 2026-09-11
- Kimi K2.8 Preview rolls out: near-K3 coding performance, 1M context for all tiers — teortaxesTex · 2026-09-11
- Looking for a classifier of software engineering task shapes to pick models per task — StewartalsopIII · 2026-09-11
- DeepSeek V4 Pro API to continue after Sept 2026, billing unchanged — teortaxesTex · 2026-09-11
- DeepSeek V4.1 Flash Hits 98% of GPT-6 Astra's Score at 1.4% of the Cost in Third-Party Benchmark — ayushtweetshere · 2026-09-11