Every 1M-Context Model Compared: Which Window Is Real

A million-token context window was a differentiator in 2025. By August 2026 it is table stakes: Kimi K3, GLM-5.2, DeepSeek V4 Pro, DeepSeek V4 Flash and MiniMax M3 all advertise it.

Which raises a more useful question than "who has the biggest window". They all have the same window. The question is whose window works.

Advertised context is a maximum, not a promise

The advertised figure is the largest input the model will accept without returning an error. It says nothing about whether the model will reliably attend to a detail buried at the 600,000-token mark.

Every model degrades across long inputs, and the degradation is not uniform. Attention tends to be strongest at the beginning and end of a prompt and weakest in the middle — a pattern well enough established to have a name. A fact placed at 5 percent depth is retrieved far more reliably than the same fact at 50 percent.

So "1M context" across five models does not mean five equivalent capabilities. It means five models that will accept the same input size and then behave differently.

How the five differ underneath

Active parameters are the best available proxy for how much reasoning the model brings to bear on that context.

• Kimi K3 — roughly 104B active. The most compute per token of the group.
• DeepSeek V4 Pro — roughly 49B active, MIT licensed.
• GLM-5.2 — roughly 40B active, MIT, plus a 128K maximum output that none of the others match.
• MiniMax M3 — natively multimodal, so the window holds images as well as text.
• DeepSeek V4 Flash — roughly 13B active, MIT, and by far the cheapest way to read a large input.

Kimi K2.6 is the notable absence at 256K, which is worth knowing if you were assuming the whole Moonshot line was 1M.

Test it yourself: the retrieval probe

You do not have to take any vendor's word for this, and the test is simple enough to run in an afternoon.

Take a document representative of your real inputs — your own code, your own logs, not a synthetic corpus. Insert a specific, unambiguous fact at 10, 25, 50, 75 and 90 percent depth. Ask the model to retrieve it. Repeat at several total lengths: 50K, 200K, 500K, 900K.

You will get a grid showing where each model stops being reliable. That grid is worth more than every published context figure combined, because it is measured on your data rather than on a benchmark designed to be passed.

Two refinements make it sharper. Use a fact the model cannot infer from context, or you are measuring guessing rather than retrieval. And test reasoning over distant facts, not just retrieval — ask it to combine two facts placed far apart, which is harder and closer to real use.

Cheaper reading changes the architecture

The most underrated consequence of the 1M window becoming standard is what it does at the budget end.

DeepSeek V4 Flash reading 500K tokens costs a few cents. That makes "load a large slice of the repository and ask a question" a viable pattern rather than an extravagance, and it competes directly with building a retrieval pipeline.

For a small or medium codebase, feeding the relevant subset to a cheap 1M model is often simpler and better than embedding, indexing and retrieving — no index to maintain, no chunking strategy to tune, no staleness. RAG vs long context covers where that stops holding.

Maximum output is the forgotten limit

Everyone checks input context. Almost nobody checks output.

Most models cap generation far below their input window, and you discover the ceiling by hitting it mid-function. GLM-5.2's 128K maximum output is unusual and matters specifically for generative work — full migrations, complete test suites, whole translated files.

If your workload produces large artefacts rather than analysing large ones, check that number before the context window.

Practical guidance

• Reading enormous inputs cheaply — DeepSeek V4 Flash.
• Generating enormous outputs — GLM-5.2, on the 128K ceiling.
• Images inside the window — MiniMax M3.
• Reasoning hardest over long context — Kimi K3, on active parameters.
• Permissive licence with depth — DeepSeek V4 Pro.

And whatever you pick, run the retrieval probe before designing a pipeline that assumes the advertised number.