Anthropic says its merges rose from about 2,500 a day in May to more than 30,000 by September. Its CI systems now do about 25 million checkouts a day. At that volume, the useful question is what a job can skip. For the other speed results in this story, it is what the reported time leaves out.
Scion Cut Checkouts From Minutes to Seconds
Anthropic built Scion for its enormous monorepos. Scion keeps history on a server, while each client keeps a bounded cache. A CI job can ask for a diff without checking out the files, and Anthropic reports cache hit rates above 99% at each of Scion’s three layers.
Anthropic says checkouts fell from minutes to seconds. It also says the infrastructure uses 200 times less CPU and memory than its old Git mirror fleet. Those are large reported gains, but Anthropic’s post does not explain how it measured the checkout and resource figures. The result is promising on its face; the measurement behind it is harder to assess.
Anthropic says Scion is for enormous monorepos, not a general Git replacement, and it will not be open-sourced. The checkout result also leaves another part of the workflow open. Anthropic tests after merge but before deploy and lets Claude fix conflicts, yet its post does not document how Scion handles two merges racing each other.
An Hour of Audio in Seconds Comes With More Errors
The developers of a Mac transcription app report processing one hour of audio in 12.6 seconds on a base 8 GB M3 MacBook Air. IBM Granite Speech took 25.7 seconds on the same machine in their test. That is a striking time for an hour of audio, within the scope of the developers’ comparison.
The app uses Google’s Embedding Gemma 2 to turn audio into vectors. A small reader moves those vectors toward written words, choosing from about 229,000 English words. A decoder removes blanks and repeats, then another small model adds punctuation and capitals.
The developers say about 5% of words are wrong, versus 2 to 3% for leading models. Model cards report 5.8 to 6.8% errors on clean English audiobook speech and about 15% on harder sets. The processing time is easy to notice, but the error figures belong beside it: a fast transcript still has to be read with its accuracy in mind.
The developers say the audio stays on the Mac. They do not document how the app splits longer files, a missing detail when reading the reported result for an hour of audio.
A Smaller Model File Does Not Settle the GPU Choice
A 27 billion parameter model normally needs about 54 GB at 16 bits per weight. Smaller weights change what fits on a local GPU. File size alone, though, does not tell you how good the answers are or how much room remains for the cache that grows during a long conversation.
The speaker’s picks make those memory limits concrete:
- 8 GB: Bonsai Heretic, with a 5.95 GB file.
- 12 GB: Bonsai V2 with multi-token prediction, at 7.66 GB.
- 16 GB: Bucoid for long context, or Zerob Server X IQ3S at 11.80 GB for a smaller file.
- 24 GB: Kleti’s Q4KM build at 16.8 GB.
Multi-token prediction can make generation faster by checking proposed tokens together. It also takes memory. Bonsai V2’s file grows from 7.21 GB to 7.66 GB with it, and Bucoid’s card says the context that fits on a 16 GB GPU falls from about 110,000 to 80,000 tokens. Bolding’s model card claims about 40% faster generation for its MTP build. These are separate model card reports, but they put the same choice in view: generation speed and room for context both matter.
Refusal results need a separate reading. In a paper’s test on 70 billion Llama 3, an edit cut refusals from about 99% to 4%, while MMLU moved from 79.9 to 79.8. The paper did not test Quen 3.8 or Gemma 4. The local builds’ refusal counts come from different releasers’ own prompt sets, so those counts cannot be compared directly.
More Activity Can Leave Work Unfinished
Each speed result answers a narrow question. Anthropic reports quicker checkouts without explaining its measurements or documenting racing merges. The Mac app’s developers report very fast transcription alongside more word errors. A local model build can promise faster generation while taking memory that could otherwise hold context.
NateBJones raises the concern that sits beyond those measurements. AI can make it easy to produce prototypes, plans, and tests for new ideas. That activity can consume the capacity AI freed up while the important work remains unfinished.