Dev
Tooling, languages, APIs and the daily craft of building software.
Claude’s 17x quantum gate cut is measured against the slowest circuit
Anthropic says Claude Mythos 5 found an elliptic-curve circuit with 17x fewer gates for 12% more qubits. We read Table S3 of the paper: the 17x is against the most...
Read moreDetailsDeepSeek’s search API caps every account at 5 requests a second
DeepSeek opened Search and Fetch on 17 October at $0.75 per 1,000 successful calls. We read the guide: one account gets 5 requests a second whatever the key, Fetch has...
Read moreDetailsOpenAI’s free prompt voices only speak inside GPT-Live
Describe a voice, get a voice id, pay nothing. We read OpenAI's custom voices guide and found where prompt voices can't go: outside GPT-Live, under Zero Data Retention, or into...
Read moreDetailsFLUX 3 Image Klein runs in 8GB, but not with the upsampler it wants
Black Forest Labs says Klein runs on 8GB RTX cards. The INT8 generator and text encoder files already total 10.69 GB, and the 9B prompt upsampler BFL recommends adds 6.72...
Read moreDetailsDeepSeek swapped deepseek-flash today, and there’s no id to pin back
On 17 October DeepSeek put V4.1-Flash-1017 behind deepseek-flash, same price, bigger agent scores. The API offers no dated id for the old build, the new weights ship without a Jinja...
Read moreDetailsGranite Switch 4.1 won’t install next to this month’s vLLM
IBM released Granite Switch 4.1 on 15 October. The model cards pin vLLM below 0.28, the plugin allows up to 0.30, and neither covers vLLM 0.31.0. Speech is English only.
Read moreDetailsBeam’s weights are out, and vLLM support is still an open pull request
Reflection published Beam 1.0 weights on 16 October: 1,002.68 GB in BF16, 293.65 GB in NVFP4, plain Apache 2.0. Transformers 5.20.1 loads it; the vLLM day-0 support is an unmerged...
Read moreDetailsIs Opus 5.5 fast mode worth $40 per million output tokens?
Since 15 October, fast mode on Claude Opus 5.5 needs no waitlist or beta header. We stacked every multiplier on the $8/$40 price and found the catch: fast and standard...
Read moreDetailsFine-tune on Muse Spark Open outputs and you owe Meta a credit line
Muse Spark Open went up on Hugging Face on 15 October under an MIT-style licence with three strings: a Built with credit that follows models trained on its outputs, a...
Read moreDetailsNemotron 3.5 Bolt says 24GB on the label and 28GB in the recipe
NVIDIA released Nemotron 3.5 Bolt 74B-A9B on 15 October. The files the recipe loads total 24.75 GB, there is no 24 GB discrete recipe, and the PinchBench #1 claim trails...
Read moreDetails