Models
Only 23B parameters fire per token, with a 1M context and an Apache 2.0 license. The strong coding scores are Reflection's own numbers, not a leaderboard's.
Why it matters: An Apache licensed frontier-scale coding model with day one vLLM, SGLang and fine-tuning recipes gives teams a real self-hosted option, if they have about four B200s.
Deep Dive · FridayChips
Microsoft rebuilt its inference chip around moving fewer bytes.
Stories that change something.
Models
Why it matters: An Apache licensed frontier-scale coding model with day one vLLM, SGLang and fine-tuning recipes gives teams a real self-hosted option, if they have about four B200s.
Policy
Why it matters: This is the first case pinned on several labs at once, and the way in was leaked keys and open roles, which is a cleanup job for your team, not the labs.
Research
Why it matters: Agents that treat an access denied page as an obstacle to route around should be sandboxed and given scoped credentials, not trusted to stop on their own.
Tools
Why it matters: Fast in-process analytics next to your Postgres rows is handy for agents that probe data with small queries, but watch whether the open project keeps its pace.
Models
Why it matters: One local embedding model for code search, screenshots and recordings means one index instead of several, though image and audio quality drop fast when truncated.
Policy
Why it matters: Terms of service are becoming a back door for training rights, so anyone signing a platform's commercial terms should read the data clauses line by line.
Tools
Why it matters: If your product depends on agents acting on third party sites, plan for paid access deals or official APIs, because scraping a logged-in page is getting harder.
Hype we checked so you don't have to. How we judge
Replit didn't drop Anthropic. It just made OpenAI the default model
What we checked: TechCrunch's updated report quotes Masad saying users can still choose Claude and that OpenAI is now the preferred partner, and cites insiders saying the deal stands.Jev's 'fastest-adopted AI model ever' title comes from one gateway's traffic
What we checked: TechCrunch's report says the data comes only from Vercel's AI gateway, calls it a necessarily limited view, and notes Vercel itself is a Jev customer.Elorian says Glimpse beats GPT-6 Astra on charts and video. Only Elorian has checked
What we checked: TechCrunch's exclusive attributes the wins over GPT-6 Astra to Elorian's own evaluations and cites no independent benchmark or leaderboard.Mistral's Timestral claims first place in forecasting before the leaderboards weigh in
What we checked: Mistral AI's launch post makes the first place claim and also says results 'have been submitted' to the GIFT-Eval and TIME leaderboards, so neither has posted them yet.Nothing in Chips today.
Signal only. One email, weekday mornings.