Skip to content

LLM & VLM

A visual guide to the shared ideas behind language and vision language models.

Updated View as Markdown

A model name tells you what to look up. An architecture tells you how information moves. This collection builds that understanding, then uses concrete models to examine what changes, why it matters, and where the tradeoffs appear.

Choose your view

Mechanism mapOne subject, three useful views

01 · Foundations

Trace tokens through attention, feature transformations and residual connections.

02 · Architectural Evolution

Compare what changes in history, position, capacity, depth and visual evidence.

03 · Model Deep Dives

Inspect a named model through its report, released artifact and implementation.

Reading paths, not a chronology or a ranking.

Figure 1. Foundations, shared design choices, and concrete models. A reading map created here, not a model lineage.

If you want to… Start here
Understand the original computation Transformer foundations
Compare the mechanisms behind different models Architectural Evolution
Inspect a report and released implementation Model Deep Dives

In Figure 1, Foundations is one complete walkthrough. Evolution connects ideas across families. Deep dives provide the report, figures, training recipe, inference contract, and evidence for a named release. Move between these views as needed.

The questions beneath the names

Start with the question that interests you:

  • What should a model remember? Compare attention and memory, from KV sharing and latent caches to recurrent state and selected history.
  • How should it locate evidence? Separate position and long context from a promise of reliable recall.
  • Where should it spend computation? Understand FFNs and experts, including the difference between feature gates and routing.
  • How should information travel through layers? Examine depth and residual paths.
  • How should it read an image or video? Trace multimodal integration from pixels to usable visual evidence.

These choices can coexist. DeepSeek-V2 combines latent attention with sparse experts; Gemma 3 combines local and global attention with a visual interface; Kimi K3 adds selective information flow across depth.123 None is adequately described by one label such as “efficient” or “multimodal.”

Architecture is one part of the explanation

Data, objectives, optimization, and inference budgets can change behavior without introducing a new neural block. DeepSeek-R1 is a useful example: its reasoning training builds on an already pretrained backbone.4 Keep the computation and the learning process separate before attributing a gain.

This is a curated learning map, not a ranking or an exhaustive release tracker. The model catalog covers named releases; the research agenda explains what evidence would improve these architectural explanations. Selected mechanisms were rechecked October 10, 2026. No model weights or benchmarks were run here.

References

  1. DeepSeek-AI. DeepSeek-V2, 2024, section 2: MLA and DeepSeekMoE. ↩

  2. Gemma Team. Gemma 3 Technical Report, 2025, architecture and visual interface. ↩

  3. Kimi Team. Kimi K3: Open Frontier Intelligence, 2026, section 2: token, channel and layer mixing. ↩

  4. DeepSeek-AI. DeepSeek-R1, 2025, section 2: reasoning training and distillation. ↩

Navigation

Type to search…

↑↓ navigate↵ selectEsc close