LLMs: Everyone Quotes Parameters, Tokens, Context Window Counts. Almost Nobody Knows What That Means.

Parameters = how much the cook has learned

Imagine hiring someone to cook in your kitchen. You want to know how good they are. Four things end up on their CV, and each one tells you something completely different.

A parameter is one number. That's it. Something like 0.0037. Not a fact, not a word, not a rule — just a dial setting, one of billions, determining how strongly one signal inside the machine pushes on the next. When people say a model has "397 billion parameters," they mean it contains 397 billion of these dials.

Think of everything the cook has ever absorbed: every recipe, every technique, every time they burned something and learned not to. Parameters are the storage holding all of it.

Two things worth knowing. No single parameter means anything — you can't open the model, find dial number 14,882,301,776, and discover it's the one that knows Paris is in France. Remove any one and nothing observable happens. Knowledge lives in the pattern across billions of dials at once, the way a photograph isn't in any single pixel. Nobody typed these numbers in either: training is the process of nudging all of them a fraction at a time, billions of times over, until sensible output starts coming out. No human designed the final values and no human can read them back.

And bigger isn't automatically better. A cook who read ten thousand cookbooks isn't necessarily better than one who read a thousand and actually practised. Parameters measure capacity, not skill. Plenty of enormous models lose to well-trained smaller ones.

"Mixture-of-experts" = a brigade, not one cook

You'll see specs written as "397B total, 17B active." That's a mixture-of-experts model, and it's the single most useful thing to understand about modern AI.

Instead of one cook who knows everything, the kitchen runs a brigade of specialists: pastry, sauces, grill, garde-manger. When an order comes in, only the two or three relevant cooks wake up. Everyone else stands idle.

So a 397-billion-parameter model might use only 17 billion of those parameters for any given question. It gets to be enormous — all that specialist knowledge available — while running at the speed of something twenty times smaller.

The catch: the whole brigade still has to be in the building. You need memory for all 397 billion even though only 17 billion work each order. Total parameters decide what hardware you need. Active parameters decide how fast it runs.

Context window = how big the countertop is

The context window is how much the cook can keep in front of them while working.

It's measured in tokens — chunks of text, roughly ¾ of a word each. "Unbelievable" is about three tokens; "the" is one. A million-token context window is around 750,000 words: eight or nine novels spread across the counter simultaneously. Hand the cook your entire company's paperwork and they can consult all of it while answering, rather than forgetting the start by the time they reach the end.

Again — capacity, not skill. A bigger countertop doesn't make a better cook. It means you can hand them bigger jobs.

Parameters live inside the model and never change while you use it. Tokens are text flowing through it, different every conversation. Context windows are measured in tokens; model size in parameters. When an AI company bills you, it bills per token — never per parameter.

License = who owns the recipe book

This one has nothing to do with ability. It's about freedom and price.

Most AI works like a restaurant: the chef stays in their kitchen, you pay per meal, and they can change the menu or raise prices whenever they like. You never see the recipes. Open weights means they publish the recipe book — you download it, run it on your own machine, no internet connection and no bill.

But "open" comes in strengths, and this is where headlines get sloppy:

Articles describing a batch of models as "MIT licensed" are usually rounding off. The differences matter enormously if you plan to build a business on one.

The benchmark = the actual exam

Here, finally, is a number about quality. SWE-bench Verified hands the model real, previously unsolved problems from actual software projects and asks it to fix them. "80.6%" means it fixed four out of five. Other exams test other subjects: AIME is competition mathematics, GPQA Diamond is PhD-level science, Terminal-Bench is operating a command line.

This is the number telling you whether the cook can actually cook. Everything else tells you how much they've learned, how much they can hold, and what they cost.

Treat these scores with suspicion, though. Almost all are self-reported by the company selling the model. And a benchmark is one specific test — DeepSeek V4-Pro scores 80.6% on SWE-bench Verified but only 8% on DeepSWE, a stricter version of roughly the same task. Same model, same skill, wildly different number, because the exam changed.

One more: how much does it weigh?

Every parameter must be physically stored, and you choose how many bits to spend on each: 16 bits (2 bytes) is full quality straight out of training; 8 bits halves the size with barely noticeable loss; 4 bits quarters it with a small but real loss.

That's quantization, and "a 4-bit quant" is the phrase you'll see. It's like re-saving a photo as a smaller JPEG — still the picture, slightly muddier. In practice a "4-bit" model averages closer to 5 bits per parameter, because the most sensitive layers get kept at higher precision. Multiply parameters by bytes and you get the number that decides whether you can run it at all.

The comparison table

Six open-weight models that shipped in the first half of 2026 — the ones everyone points at when arguing that free AI has caught the paid kind.

Two notes before reading anything into this. GLM-5.2's 62.1% is on SWE-bench Pro, a harder exam — not comparable to the others in that column, and it does not mean GLM is worse. And the four scores that are comparable sit between 76.4% and 80.6%, a spread narrow enough to be mostly noise. Anyone telling you a 0.4-point gap makes one model the winner is selling something.

So which one should you use?

These models are far more alike than the marketing suggests. The real differences aren't in the exam scores — they're in the other four columns.

Building a product? Qwen 3.5. The only one here under Apache 2.0, which includes an explicit patent grant. Every other license on this list is either a modified variant or a custom document your lawyer has never seen. At 397B/17B it's also the cheapest to run of the genuinely capable options, and it covers 201 languages — which matters enormously if your users aren't English speakers. Its 76.4% is the lowest comparable score here, and that gap will not be what limits your product.

Want the strongest thing available, cost no object? DeepSeek V4-Pro. Top score, a million tokens of context, clean MIT license. It also needs roughly a terabyte of memory even compressed, so "self-hosting" means a rack, not a desk. For most people this is an API model that happens to be open — the freedom is theoretical.

Want an AI that works unattended for hours? Kimi K2.6 or GLM-5.2. Both are built for long-horizon agent work — describe a goal, come back later. Kimi coordinates specialist sub-agents across thousands of steps; GLM-5.2 was explicitly tuned for self-hosted coding agents and holds a million tokens while doing it. Kimi's license has strings; GLM's is clean MIT. If you're self-hosting, that tips it to GLM.

Work involves screenshots, images or video? MiniMax M3. The only one here trained natively multimodal, so it can look at a picture of an interface rather than just read about it. Turning a design into working code, or automating something that only exists on screen, is its lane. The custom license is the price of admission — read it before you commit.

Feeding it enormous documents? Anything with 1 M context — that's every model here except Kimi K2.6. Whole contracts, entire codebases, a year of correspondence, all visible at once.

Actually want to run one on your own computer? None of these. This is the part the excited articles skip. The smallest model in the table needs roughly 178 GB of memory after aggressive compression. A well-specified laptop has 32 GB. A €500 gaming graphics card has 16 GB. These are "free" the way a shipping container of flour is free if someone gives you one — the gift is real, and you still have nowhere to put it. What you can genuinely run at home are the smaller siblings of these families: 8B to 70B versions that fit in 6–48 GB and are, for ordinary use, remarkably good. Start there. The giants in this table are for people with server racks or a cloud account.

What none of these numbers tell you

The spec sheet is silent on almost everything determining whether you'll like using a model: whether its writing has any life in it, whether it admits uncertainty or bluffs, how it behaves on the tenth turn of a hard conversation, whether it quietly ignores half your instructions.

There's no column for that. There probably can't be. The only test that counts is putting the same real task to two of them and seeing which answer you'd actually use.

Which, if you think about it, is exactly how you'd hire a cook. You wouldn't read their CV. You'd ask them to make you dinner.

Model specifications compiled August 2026 from vendor documentation and independent benchmark aggregators. Benchmark scores are largely vendor-reported and should be treated as claims rather than measurements. Memory figures assume ~5 bits per parameter — the realistic average for a "4-bit" quantization — and exclude the additional memory needed for context, which at a million tokens is substantial.