The Four Axes of Modern LLMs

The compass

Every modern LLM can be located on four independent axes. Independent meaning a change on one axis is mostly orthogonal to the others — DeepSeek-R1 and DeepSeek-V3 share the same backbone, the same inference engine, and live in roughly the same product surface, yet they behave very differently. The difference is purely on the post-training axis.

Axis 1 — Backbone architecture

How is the network actually wired? Dense vs MoE vs hybrid. How many parameters exist, and how many activate per token? What attention variant — full attention, GQA, MQA, MLA? How is position encoded — RoPE, ALiBi, NoPE, YaRN-extended? Anything you can read from the config.json in a Hugging Face repo lives here.

Axis 2 — Post-training

What was done to the pretrained base after the autoregressive crawl-the-internet phase? RLHF, DPO, GRPO, RLVR, distillation, reasoning-focused RL. This is where raw next-token machines become assistants, and where most of the "personality" and capability of a 2025–2026 frontier model actually comes from. The DeepSeek-V3 → DeepSeek-R1 jump is pure post-training; same weights file shape, different training recipe.

Axis 3 — Inference strategy

How is the model invoked at runtime? Standard autoregressive, extended thinking with reasoning tokens, test-time compute scaling, speculative decoding, beam search, structured output via grammar-constrained decoding. Same weights, different runtime behavior. Claude Sonnet with extended thinking off and Claude Sonnet with extended thinking on use the same checkpoint — only the inference axis changes.

Axis 4 — Product behavior

What scaffolding wraps the model? RAG, tool use, agent loops, system prompts, guardrails, memory, multi-step orchestration. ChatGPT-the-product is a thin layer over o3-the-model, and "Claude with Computer Use" is the same Claude with a different tool surface. Application layers, not architecture.

Code

Reading a model card through the four axes·python

# Sample card text:
# "DeepSeek-R1, a 671B-A37B MoE model with MLA attention,
#  trained with GRPO reinforcement learning on top of V3,
#  served with extended thinking <think> blocks,
#  available with native tool use."

axes = {
    "backbone":       "671B-A37B MoE, MLA attention",   # axis 1
    "post_training":  "GRPO RL on top of V3",            # axis 2
    "inference":      "extended thinking <think> blocks",# axis 3
    "product":        "native tool use",                 # axis 4
}
# Each phrase belongs to exactly ONE axis — that is the test.

The same weights, different axes·python

# The same Claude Sonnet checkpoint serves both:
fast_response   = call_claude(messages, thinking=False)
careful_response = call_claude(messages, thinking=True, budget=8000)
# Backbone, post-training, product — all identical.
# Only the inference axis (axis 3) changed.

Exercise

Pick a model you actually use (Claude, GPT, Gemini, Llama, Qwen, anything). Make a 4-row table — one row per axis — and fill in what you know about that model on each axis. The point is to notice which axis is the *vaguest* in your mental model — that is exactly where the marketing has the most room to lie.