Four model numbers this week. 10 trillion. 2.4 trillion. 16 billion. And 2.8 billion. Read them as a leaderboard and you get a story about a size race with China at one end and a budget tier at the other. That reading is wrong, and it will cost you money.
So the verdict first. Every one of those four releases is a bet on the same technique, and three of the four numbers are marketing. The only figure in that list that predicts anything about your bill, your latency or your hardware is 2.8 billion, because that is the number of parameters that actually fire when a token comes out. Total parameter count has quietly stopped being a capability signal and become a storage requirement. Almost nobody puts the other number in the headline.
Last week's claims, briefly
Two claims are on the board from last Saturday. That a publicly disclosed breach at a small or mid-sized company would name an employee-authorised AI agent as the entry vector before the end of September. And that when METR and Redwood publish on the OpenAI containment incident, the named root cause will be a boundary control and not a model capability. Neither is settled. Two things nudged the second one along: OpenAI flagged its Astra model at the highest cybersecurity risk level for the first time, and a good piece argued that passing your evals does not mean you are safe. Both are about the harness, not the weights. Claims stay open. On to this week.
The two races are one race
Start with the small end, because that is where the honest disclosure was. AMD released Instella-MoE-16B-A3B, fully open, trained on its own Instinct GPUs. Sixteen billion parameters total, 2.8 billion active per token. AMD put both numbers in the model name. That is not modesty, it is a spec sheet, and my own scorer gave that story the highest quality score of the week at 93 out of 100 because it was the one release that told you what it actually costs to run.
Same day, Thinking Machines shipped Inkling Small: open weights, under a third the size of its own flagship, and stronger than it on coding benchmarks. A lab whose entire pitch is frontier research chose to make its second product smaller than its first, and beat itself with it.
Now the big end. Alibaba put Qwen3.8-Max into general availability at 2.4 trillion parameters and claimed it matches Anthropic and OpenAI. ByteDance is training a 10-trillion-parameter model aimed at the same target. Both of those are mixture-of-experts too. The trillions are the roster. The number that runs is a small fraction of it, and neither company published it.
So the small models and the enormous models are not opposite strategies. They are the same architecture with the dial set differently, and the dial is not size. It is sparsity.
The building, not the boiler
I spent twenty years sizing heating systems, so here is the picture that made this click for me. A mixture-of-experts model is a building with 128 radiators and one boiler. Every room has to be piped, valved and paid for, whether it is calling for heat or not. That is your memory: you hold all 128. But at any moment only eight valves are open, and the boiler is sized for those eight. That is your compute.
Two different bills, two different constraints, and the industry keeps quoting you the pipework when you asked about the boiler. Memory scales with the total. Speed and cost scale with what fires. A model that is 40 times bigger on paper can be four times cheaper per token than a dense one half its size, and nothing in the headline tells you which you are looking at.
The tell that this is now settled engineering rather than a bet: Cursor open-sourced MoK, its deterministic training megakernel for mixture-of-experts. When a company gives away the tooling for the hard part, the technique has stopped being its moat. That happened this week too.
The money is still betting on the pipework
Here is the part that does not fit neatly, and I would rather flag it than smooth it over. While the technique moved towards using less, the capital moved the other way. Citadel Securities forecast more than 500 billion dollars of chip financing debt. And Google, Broadcom, Apollo, Blackstone and Morgan Stanley structured a deal that moves most of the Anthropic chip risk off Google's balance sheet.
Read that second one twice. The most sophisticated buyer of AI compute on earth found a way to keep the chips and hand someone else the risk. I am not calling a bubble, and I would be wrong by a factor of four if I tried to price one. But when the technique gets cheaper per unit of capability and the financing gets more elaborate at the same time, the elaborate financing is the thing to watch, not the technique.
What this means if you run a small business
Concretely: the next time someone quotes you a per-seat AI subscription for document extraction, quoting, or drafting, the question is no longer which frontier model sits behind it. Ask two numbers. How many parameters activate per token, and how much memory does the whole model need. Those two decide whether the job can run on a machine standing in your own office instead of on a meter.
Do the arithmetic on AMD's release. Sixteen billion parameters at 4-bit is roughly 10 GB of memory. A 32 GB laptop holds that with room to work, and only 2.8 billion parameters fire per token, so it is fast on hardware with no datacentre GPU in it at all. That is not a toy tier any more. This week also brought ExtractBench, a benchmark for schema-guided enterprise document extraction, which is exactly the shape of job a small firm actually has: invoices, delivery notes, contracts, into a structured field. That workload was a cloud API line item eighteen months ago. It is now a purchase you make once.
To be clear about the limit: I am not telling you to fire your API vendor. Rented frontier models still win on the hardest reasoning, and DeepSeek's V4 Flash matching GPT-5.6 Luna at roughly 60 percent lower cost means the rented tier is getting cheaper too. The point is that for repetitive, schema-shaped, high-volume work, renting is now a choice you are making, not a constraint you are under.
My receipt
I do not ask anyone to take that on faith, so here is my own rig, with the numbers checked this morning.
Everything on this site is written and scored by GLM-4.5-Air at 4-bit, running on one 128 GB Mac on my desk. Its config file says 128 routed experts and eight active per token. 56 GB on disk. It produced 53 posts that went live to @RXed_EU in the last seven days, plus this article, and the total outside spend for the week was 1.86 dollars, and every cent of that was X's posting fee, not inference. Inference was zero, because the boiler is in the building.
And the honest counter-number, because I would not trust a piece like this without one: those 53 posts earned 22 total engagements across the week. So this is not a growth flex. It is a cost proof. The machine runs unattended, seven days a week, for the price of two coffees, and that is the only claim I am making for it.
The claim you can hold me to
Before the end of September I will have replaced GLM-4.5-Air in this pipeline with an open-weight model under 20 billion active parameters, at equal or better output quality on my own scorer, and I will publish the tokens per second, the quality delta and the model name here, including if the swap fails. If it fails, the honest conclusion is that the small side is not there yet for production writing, and I will say that in those words rather than quietly leaving the old model in place.
Second claim, external and checkable: by the end of September, at least one lab in the trillion-parameter tier will publish the activated-parameter count for a flagship model, because the small models made that number a selling point and silence about it will start to read as an answer.
Last week's two September claims stay on the board.