A local inference engine for DeepSeek V4 Flash hit Hacker News again today, credited to the creator of Redis. The link actually goes to a fan-built explainer site, and the live argument in the comments over whether its aggressive quantization holds up cites a benchmark as proof. The benchmark never mentions the project, and its own author says the two runs in it canβt answer that question.
A blog post making the rounds says Western labs quietly copied DeepSeekβs cache trick, and offers two price cuts as proof. The 437x figure is real arithmetic from DeepSeekβs own paper. The proof is a different matter: both price cuts also fit a simpler change on the price sheets, and neither labβs launch post mentions KV caches or DeepSeek.
Anthropicβs new prompting guide for Claude Opus 5.5 says the quiet part plainly: effort level names donβt mean the same amount of thinking from one model to the next. The guide backs its advice with nine findings from Anthropicβs own testing. Eight give a direction, one gives a size, and none gives a sample. A sibling page on the same site shows what the missing numbers look like, and they complicate the advice.
A headline said Xiaomiβs trillion-parameter MiMo-V2.6-Pro was trained for $3M. Xiaomiβs own report says something narrower: $2.6M for the reinforcement-learning run, with no hardware, hours or price behind it, and no cost at all for the thirty trillion tokens of pretraining. That narrower number is still the interesting one. The final stage of training now costs about nine times what DeepSeek spent on R1βs entire reasoning training.
Anthropic says Opus 5.5 costs 40% less than Opus 5. OpenAI says GPT-6 Sol costs 50% less than its predecessor. Both headlines depend on a number neither company prints for you: how much of your bill is cache reads. On a measured coding-agent mix the answers are about 44% and, against each other, about 29%.
TypeSafeβs Jev is billed as 193.6 times faster than frontier models. Outside tests measure 5 to 81 times faster. The gap is which rival sits in the denominator.
Z.ai says it runs all production inference for one of its models on more than 100,000 Chinese-made AI accelerators, at a per-token cost comparable to NVIDIA. The post names the quantization scheme, the Triton precision flag, and the exact library version whose C++ bindings held the Python GIL. It never names the accelerator. The most obscure claim in it is the only one I could fully verify.
Moonshotβs Kimi K3 is the largest disclosed language model on the market, at 2.8 trillion parameters. Nobody outside OpenAI, Anthropic, or Google knows how big GPT-6 Astra, Claude, or Gemini actually are, because none of the three labs discloses a parameter count. Meanwhile Hugging Faceβs own numbers say 83 percent of all downloads on its Hub go to models under a billion parameters, and only 1 percent go above a hundred billion. Three different questions, three different answers, and none of them is βhow big is AI.β
DeepSeek released V4.1-Flash this morning. The cardβs first sentence says 552B parameters. Three paragraphs later it discloses another 196B of memory. The checkpoint on disk holds 763.2B, and the model touches 8B of them to read a token. Every one of those numbers is correct, and the industry still prints one column.
Two preprints went up a day apart this month and were read as proof that the load-balancing loss fills mixture-of-experts models with redundant experts you can delete for free. One of them says the opposite in plain type. The homogenization is real, it was measured in a controlled run back in May, and it lands on the router rather than the experts. In that experiment the loss whose entire job is balancing load balanced it six times worse than not having the loss at all.
Model collapse is real, reproducible, and provable in the limit β but every canonical demonstration of it replaces the human data with machine output. Change that one assumption and the theory changes with it: accumulate real and synthetic data and the test error is provably bounded; filter synthetic data through a verifier and it beats human data outright. What degrades in every regime, including the successful ones, is diversity.
On-device inference is not compute-limited β it is memory-bandwidth-limited, and the NPU spec sheet measures the wrong thing. The accuracy curve is flat to 8-bit, gentle at 4-bit weights, and a cliff below that. And at identical 4-bit precision, the choice of runtime changes energy per token by tenfold.
page 2 of 2 · 27 essays on Models, inference & pricing