C.W.K.
Stream
Lesson 04 of 05 · published

Architecture Decision Table

~12 min · decision-table, summary

Level 0Observer
0 XP0/50 lessons0/14 achievements
0/100 XP to next level100 XP to go0% complete

The pin-this-on-your-wall table

Below is the practical comparison table for 2026 production decisions. Print it, pin it next to your screen, and consult it before making any architecture commitment.

ArchitectureCore ideaMain advantageMain weaknessBest use caseStatus (2026)
TransformerFull pairwise attentionBest recall, most mature ecosystemO(n²) cost, KV-cache growthGeneral NLP under 32K, frontier reasoningUse Now
Efficient Transformer (FA3 + GQA + sliding window)Standard attention with engineering winsPragmatic, fully ecosystem-compatibleStill fundamentally quadratic above the windowUp to 64K tokens, recall importantUse Now
Mamba / SSMSelective state-space compressionLinear time, O(1) inference memoryWeak recall, narrow LR windowThroughput- or memory-bound long contextWatch Closely
RWKVRNN-Transformer dual formConstant per-token cost, 1.5B-device provenSmaller ecosystem, recall limit ~28KOn-device, streaming, edgeWatch Closely
RetNetExponential decay retentionThree computation modesData-independent decay limits expressivenessResearch foundation, conceptual influenceResearch
HyenaImplicit long convolutions100× faster than attention at 64K+Weak recall, niche language qualityGenomics, byte-level, very long sequencesWatch (domain-specific)
Hybrid SSM-AttentionSSM majority + sparse attention layersBest quality–efficiency balanceArchitectural complexity to design from scratchLong-context production (32K–1M)Use Now
Linear Attention (Kimi Linear, MHLA)Smarter linear-attention with state structureDrop-in for Transformer, large speedupsNewer, less battle-testedLong-context summarization, streamingWatch Closely

How to read this table

Use Now means: production-ready, mature tooling, defensible to your team and your CFO. Watch Closely means: real production deployments exist, the architecture is solidifying, you should be able to evaluate it in 2026 if your workload pushes you there. Research means: conceptually important and worth understanding, but no production case in 2026 where you'd pick it over a competing alternative.

External links

Exercise

Print this table (or recreate it in your engineering notebook). For your team's three most important production workloads, write a one-line answer next to the table for each: which architecture lane does this workload live in? Refer back to the answer the next time someone proposes an architecture change without justification — the table is the conversation starter.

Progress

Progress is local-only — sign in to sync across devices.
Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)
Sign inPlease sign in to comment.

No comments yet — be the first.