Korva is a scroll of the internet’s best ideas.
On September 9, DeepSeek released 4.1 Flash, priced at $1.20 per million output tokens when DeepSeek hosts it. On September 17, Prism ML released Bonsai2, a fork of Alibaba's Qwen with 27 billion parameters, 98 percent of the big Qwen model's performance, and a size of 5.9 gigabytes that runs on a local computer with an Nvidia chip or a good Mac Studio.
Alibaba released version 2.1 of Qwen on September 20, an open weights image model that outperforms Google's Nano Banana 2 and can run at home for free. Xiaomi released Mimo on September 22, a 309 billion parameter open weight model, with Mimo Pro on par with Claude's Opus 5 across most benchmarks.
On the closed side, xAI released Grok 4.7 on September 21, and Meta released Muse and Anthropic released a new model on September 22. OpenAI released Sol and Luna on September 23.
How does a big model fit on a desktop?
A model's memory footprint is roughly its number of parameters times the bits stored for each one. Quantization shrinks the stored precision, for example from 16 bits per weight down to 8, 4, or fewer. Lower precision costs some accuracy, but careful methods keep much of the quality. The result is a file small enough to load into the memory of a consumer graphics card or a Mac.
Does a bigger parameter count mean better?
Researchers at OpenAI reported in 2020 that language model loss falls predictably as parameters, data, and compute grow. DeepMind's 2022 Chinchilla paper then showed that many large models were undertrained, and that a smaller model trained on more data could match a bigger one. Training data quality and design choices also matter. Chinchilla, at 70 billion parameters, outperformed the 280 billion parameter Gopher on many tasks.
How do benchmarks compare models?
A benchmark is a fixed set of tasks that is scored automatically, so any model can be measured the same way. MMLU, a multiple-choice test of academic knowledge introduced by Dan Hendrycks and colleagues in 2020, became a standard example. Once top models score near the ceiling, a benchmark stops separating them. Harder successors, such as MMLU-Pro, followed.
More from All-In with Chamath, Jason, Sacks & Friedberg
