AI Models: Capability Analysis, Architectures, and Family Guides
The model itself, analyzed: families, architectures, tradeoffs.
Choosing a model is a different problem from running one. The runtimes, GPUs, and serving stacks answer “how do I run this”; this cluster answers “which model, and why this one.”
The pages here analyze models as artifacts: how a family evolves generation to generation, what a model architecture changes about what the model can do, and where the quantization and capability tradeoffs sit. Every claim is anchored to vendor model cards, independent benchmarks, or measured runs — forecast values are labelled as forecasts.
Where AI Models fits in the stack
The site splits the model question into three layers:
| Layer | What it covers | Cluster |
|---|---|---|
| The model | capability, architecture, family evolution, quantization | You are here |
| The serving | runtimes, GPU fit, throughput, context behavior | LLM Hosting · LLM Performance |
| The system | routing, memory, agents, verification | AI Systems · LLM Architecture |
A capability claim made here (what a 27B dense model holds at 4-bit, what a one-forward-pass decider emits) becomes an input to the hosting and performance decisions made elsewhere. The clusters are deliberately separated so benchmark tables do not migrate into architecture essays and back.
Cluster map
The cluster is young and grows by density, not by schedule. What exists today:
- You are here — this pillar: what the cluster covers and where its boundaries sit
- Architectures — What Comes After LLMs? Mamba, Diffusion & World Models — the post-transformer landscape: state space models, diffusion language models, and JEPA world models, with the evidence for each
- Families — Qwen 4: what’s confirmed and what it means for local AI — a family reference page built around the generational 27B benchmark ladder, the Flash-Next architecture preview, and consumer-GPU fit
Planned subclusters, seeded when enough related pages exist: benchmarks (cross-model comparisons), quantization (bit-ladder references), embeddings and rerankers (retrieval-side models), comparisons (family vs family).
Model architectures
Architecture analysis asks what a design change does to capability, memory, and inference behavior — not which model tops a leaderboard this week. The anchor page covers the candidates lined up to succeed or supplement the transformer:
- State space models (Mamba) — linear scaling with sequence length through a compressed recurrent state
- Diffusion language models — parallel token generation instead of left-to-right decoding
- World models and JEPA — prediction over representations rather than tokens
The hybrid pattern — transformer attention where it earns its cost, linear attention where it does not — is already in shipping models, and What Comes After LLMs? traces both the research line and the production evidence.
Model families
Family pages track one model line across generations: what each release changed, which benchmark gains are real versus vendor-reported, and what the next generation means for hardware you may already own.
The first one covers the most-run open-weight family on consumer GPUs:
- Qwen 4: what’s confirmed and what it means for local AI — Alibaba’s confirmed roadmap, the Qwen3.5 → 3.6 → 3.8 27B benchmark ladder, Flash-Next as a runnable architecture preview, and the memory targets a Qwen 4 27B will land at.
What does not belong here
Boundary rules keep the cluster honest:
- Serving and tuning questions — how fast a model runs, how to fit context, which runtime — belong in LLM Performance and LLM Hosting. A family page links to those measurements rather than reproducing tok/s tables.
- Using models inside a system — routing tiers, memory, agent design — belongs in AI Systems and LLM Architecture.
- Retrieval-side models — embeddings and rerankers are analyzed here as models, but their application inside pipelines lives in RAG.
How this cluster connects to the others
- LLM Performance in 2026 — the measured counterparts to the capability claims here: the efficient frontier of open models and the 16 GB benchmark pages turn “this model is capable” into “this model fits and runs at N t/s”.
- LLM Hosting in 2026 — the runtime and infrastructure decision that follows model choice.
- What Comes After LLMs? and the family pages cross-reference each other: architecture previews in a family (Flash-Next) are the near-term evidence for the architectural shifts the landscape page describes.
