The 2026 LLM Reading Map: A Practical Order for Working Through the Whole Model Landscape

Fifteen model families, three licensing categories, and at least six distinct architectural ideas (MoE, MLA, KDA, DeltaNet, Mamba/LatentMoE, sparse attention) is a lot to hold in your head at once. This post is the map, not the territory — a condensed way to navigate the full 2026 model landscape without reading fifteen separate briefings cold, plus the one terminology distinction (closed vs. open-weight vs. open-source) that trips up more procurement and architecture conversations than any benchmark score ever does.

1. The terminology that actually matters​

Before any model comparison, get this distinction precise — it changes procurement conversations, legal review scope, and architecture decisions more than any benchmark number:

Category	What it means	Can you self-host?	Can you inspect/modify weights?	Example from this series
Closed source	Model accessible only via a vendor's hosted API; weights never released	No	No	GPT-5.6, Claude 5, Gemini 3.x, Grok 4.5, Nova
Open-weight	Weights are downloadable and runnable, but under a custom or restricted licence (usage limits, field-of-use restrictions, scale thresholds)	Yes	Yes, within licence terms	Kimi K3 (custom licence), Llama 4 (community licence, 700M MAU threshold)
Open source (OSI-approved)	Weights downloadable under a genuinely unrestricted licence (Apache 2.0, MIT) with no field-of-use or scale restrictions	Yes	Yes, unrestricted	Qwen 3.5, Gemma 4, Mistral Small 4, GLM-5, Command A+, gpt-oss

The practical consequence: "open-weight" and "open source" get used interchangeably in casual conversation, and that's exactly where legal and procurement risk hides. A custom-licensed open-weight model can still have real usage restrictions — Kimi K3's licence terms, Llama 4's monthly-active-user threshold — that a genuinely OSI-approved Apache 2.0 model (Qwen 3.5, Gemma 4, Mistral Small 4, GLM-5, Command A+, gpt-oss) simply doesn't carry. Always check the actual licence text for the specific model and version you're deploying, not a category label.

DeepSeek and Nemotron 3 also ship open weights under release-specific terms (DeepSeek's licence and NVIDIA's open-model licence for checkpoints, datasets and recipes can differ by artefact) — check the text attached to the exact download you use.

2. The closed-source platforms​

Model family	Positioning	Briefing
GPT-5.6 (Sol/Terra/Luna)	Durable capability tiers, routed reasoning, agentic ultra mode	Read the briefing
Claude 5 (Fable/Opus 5/Sonnet 5)	Constitutional AI, native MCP, Opus as new daily-driver default	Read the briefing
Gemini 3.x (Flash/Pro/Live)	Platform breadth — grounding, Workspace, real-time voice	Read the briefing
Grok 4.5	Cursor-co-trained coding/agent model, RL at scale	Read the briefing
Amazon Nova (+ Bedrock, Act, Sonic)	AWS-native integration, sovereignty within AWS, Nova Forge customisation	Read the briefing

3. The open-weight and open-source families​

Model family	Architecture	Licence	Briefing
Kimi K3	2.8T/104B-active MoE, KDA hybrid attention, LatentMoE	Custom (open-weight)	Read the briefing
Qwen 3.5	397B/17B-active hybrid DeltaNet+MoE	Apache 2.0	Read the briefing
Gemma 4	E2B–31B, local multimodal	Apache 2.0	Read the briefing
Llama 4 (Maverick/Scout)	MoE	Community licence, 700M MAU threshold	Read the briefing
DeepSeek (V3/R1/V4)	MLA, MoE, GRPO-trained reasoning	Open weights (check release)	Read the briefing
Mistral Small 4	119B/6B unified chat+instruct+code	Apache 2.0	Read the briefing
GLM-5	Sparse attention, agentic-engineering-first training	Apache 2.0	Read the briefing
Nemotron 3 (Super/Ultra)	Super 120B/12B-active, Ultra 550B/55B-active, hybrid Mamba+LatentMoE, open recipes	NVIDIA open-model licence (check artefact)	Read the briefing
Command A+	218B/25B, sovereign AI / enterprise focus	Apache 2.0	Read the briefing
gpt-oss (120b/20b)	MXFP4-quantised reasoning models	Apache 2.0	Read the briefing

4. Recommended study order​

If you're building context from scratch rather than jumping straight to the model relevant to your current decision, this order builds concepts progressively rather than repeating them:

1. Start with the terminology table above — get closed vs. open-weight vs. open-source precise before anything else, since it recurs in every subsequent briefing.
2. GPT-5.6 and Claude 5 — the two closed frontier families most teams compare first; establishes the tiered-pricing mental model used across the whole series.
3. Gemini 3.x and Grok 4.5 — rounds out the closed-frontier picture with a platform play and a specialised coding/agent play.
4. Amazon Nova — introduces the "integration and governance over raw capability" decision frame, relevant again later for Command A+.
5. DeepSeek — read this before Kimi K3 and Qwen 3.5; it introduces MLA, MoE, and GRPO, the foundational ideas that recur (in variant form) across most of the open-weight series.
6. Kimi K3 and Qwen 3.5 — the two largest open-weight frontier contenders, and the clearest illustration of the open-weight vs. open-source licensing distinction (custom licence vs. Apache 2.0) side by side.
7. Llama 4 — the licensing deep dive; read this once you understand why licence text matters, so the 700M MAU threshold lands with full context.
8. Gemma 4 and Mistral Small 4 — the deployment-flexibility and consolidation arguments, most relevant once you're planning an actual self-hosted architecture rather than just comparing capability.
9. GLM-5 and Nemotron 3 — the two most architecturally specialised entries (agentic-first training, Mamba hybrid + open recipes); read after you have the MoE/attention-variant vocabulary from steps 5–6.
10. Command A+ and gpt-oss — close with these two: Command A+ crystallises the sovereignty argument, and gpt-oss is the single most surprising licensing fact in the whole series (Apache 2.0, from OpenAI) — a good note to end on.

5. How to use these briefings versus the model reference pages​

Each blog post in this series is an opinionated engineering briefing — when to pick this family, the traps that catch teams in practice, cost intuition, and how to evaluate the vendor's own launch claims. They are deliberately not exhaustive architecture references. For the detailed architecture, API surface, and specification-level reference material behind each family, use the AI model landscape 2026 overview and the individual model reference pages linked from each briefing's "Reading order" section.

Use the briefings to decide; use the reference pages to implement.

Verdict​

Fifteen models across five closed platforms and ten open-weight or open-source families is genuinely a lot of surface area, but the underlying decision structure is smaller than it looks: pin your architecture to a tier or licence category rather than a model name, know precisely whether you're holding open-weight or open-source before a legal review surfaces the difference for you, and validate every vendor capability claim on your own task before it enters a cost model. Start with the terminology table, follow the study order if you're building context from scratch, or jump straight to the briefing relevant to your current decision.