Models
About 49B parameters are active, with a 1M token context, FP8 and NVFP4 builds. Mistral says it leads open models on coding; its own index chart shows a tie.
Why it matters: A frontier sized model you can host and fine-tune changes the build or buy math for coding and security agents, but the license limits model-as-a-service resale and the NVFP4 build needs Blackwell GPUs.
Deep Dive · FridayModels
SpaceXSI says it's the voice model behind Starlink's support lines and Tesla's in-car assistant, and SpaceXSI sells Grok Voice 3.0 to anyone at $0.08 a minute.
Stories that change something.
Models
Why it matters: A frontier sized model you can host and fine-tune changes the build or buy math for coding and security agents, but the license limits model-as-a-service resale and the NVFP4 build needs Blackwell GPUs.
Research
Why it matters: It is a working answer to the agent era review bottleneck: instead of another AI reviewer, generate executable tests that prove a behavior change and only page a human when the evidence is strong.
Tools
Why it matters: Generated artifacts no longer need a round trip to another app, so teams drafting specs or configs in Gemini can iterate in place, which also means more edited output to review before it ships.
Tools
Why it matters: If your voice agent runs on Pipecat, its steward is now a model vendor. The vendor neutral pledge is worth holding them to, so keep your STT and TTS providers swappable.
Tools
Why it matters: Agent driven browser tests have been too slow and costly for big regression suites. Caching the resolved actions is a pattern worth copying even if you never adopt the library.
Tools
Why it matters: iOS and macOS developers can pick an open on-device model for languages Apple handles poorly without rewriting their pipeline, and keep audio off the network.
Models
Why it matters: Versioning a model like a library, with patch releases you compose on one endpoint, is a cleaner upgrade path than redeploying a whole checkpoint every time a capability improves.
Policy
Why it matters: The line it draws, training as fair use but verbatim output as infringement, is the distinction that would shape how you filter and attribute model output if courts adopt it.
Hype we checked so you don't have to. How we judge
Vera Rubin delivers 30x more agent throughput per megawatt? NVIDIA measured that itself
What we checked: The NVIDIA Developer blog states the Vera Rubin NVL72 results were measured by NVIDIA on the AgentX workload and are pending SemiAnalysis review, and the 30x figure is an up to number at 160 tokens per second per user. The post itself is dated Aug 24, 2026.Anyone can now get a religious exemption from using AI at work? One employer granted one
What we checked: TechCrunch reports the university reviews religious accommodation requests case by case and says this one may be reviewed again as jobs change. No court or agency ruling is cited, so it sets no precedent for other employers.AWS is headed for a trillion dollars a year? That rests on growth Jassy has only forecast
What we checked: TechCrunch quotes Jassy on the earnings call giving a $184 billion run rate, a belief that AWS can grow north of 40% in 2027 and 2028, and a five to ten year horizon for $1 trillion, alongside about $220 billion of 2026 capex. It is a projection, not a result.Signal only. One email, weekday mornings.