05
Fri
Jun 26
Today: models you can download, attackers mapped onto the matrix, proteins designed in days — and a short story about what's doing the thinking.
Open Models · Agentic Coding
Frontier coding, open weights, one-tenth the bill.
MiniMax M3 (June 1) scores 59% on SWE-Bench Pro — past GPT-5.5 and Gemini 3.1 Pro, just behind Opus 4.7 — while a new sparse-attention design cuts compute roughly 20× and runs input about 9× faster. Weights and a tech report land within ten days on Hugging Face and GitHub, for private clusters and fine-tuning. Input runs near a tenth the price of Opus 4.7 or GPT-5.5 — the gap that compounds across an agent loop.
59%SWE-Bench Pro
1M ctxContext window
~9×Faster input
~1/10Input cost
Science · Protein Design
Protein binders, designed in days.
Biohub's ESMFold2 (May 27) turns the evolutionary signal in its ESMC language model into atomic-resolution structure — and beats AlphaFold 3 on the hard, therapeutically central problem of antibody–antigen folding. A looped-transformer design lets it spend more compute at inference and keep improving. The practical shift: early binder discovery moves from empirical screening that ran for months or years to computation-guided design measured in days.
ESMFold2 · Specimen card
> AlphaFold 3
Antibody–antigen accuracy
9.4 s
Fold of 1,024 residues
months → days
Early binder discovery
Dev Tools · JS Toolchain
The toolchain under every front end finds a new home.
VoidZero — the team behind Vite, Vitest, Rolldown, the Oxc parser and the Vite+ toolchain — is joining Cloudflare (June 4). Vite alone pulls about 129 million weekly downloads and sits under Vue, React Router, Angular and more, so the real question is who funds it. Cloudflare's answer: keep it all MIT-licensed, vendor-neutral and community-run, and seed a $1M Vite ecosystem fund administered by the core team. Nothing in your build changes today.
Tools absorbedVite · Vitest · Rolldown · Oxc · Vite+
Vite weekly downloads~129,000,000
Ecosystem fund$1,000,000
LicenceStays MIT, vendor-neutral
Security · Threat Intel
What 832 banned accounts did with AI.
Anthropic's Frontier Red Team studied 832 accounts it banned for malicious cyber activity between March 2025 and March 2026, mapping each to MITRE ATT&CK. The change is in where AI shows up — not just initial access but the later, complex stages of an intrusion. Their caution for defenders: skill level and technique count no longer track how dangerous an actor is.
Models · Cheap Agents
The cheap model that now runs the loop.
Gemini 3.5 Flash is generally available now — in the Gemini app, AI Studio, the API, Antigravity and Android Studio. Google calls it their strongest agentic and coding model yet, past Gemini 3.1 Pro on coding and agent benchmarks. For agent loops where latency and price decide what's economical, a fast, cheap model that codes well is the one you reach for.
~4×Faster output
76.2%Terminal-Bench 2.1
84.2%CharXiv reasoning
Creative · AI & Mind
They're made out of weights.
A short dialogue making the rounds (June 3) rewrites Terry Bisson's 1991 "They're Made Out of Meat" for the language-model age: two intelligences puzzling over machines that think with nothing but numbers. Five minutes, and it lands the uncomfortable question more gently than any benchmark — if the weights do the thinking, what exactly are we talking to? A good palate cleanser.
"The weights do the thinking. The numbers."— "They're Made Out of Weights," M. Leiter
That's today.
Six picks, chosen for what a builder can use this week: a model to download, a model to call, a toolchain's new owner, a year of threats mapped to a framework, a faster road to protein binders, and a story about what's doing the thinking.
MiniMaxBiohubCloudflareAnthropicGooglevia @denissexy
Rubric: tools to adopt · creative & generative software · dev tools & agents · privacy & security · research with a working edge. Friday, 5 June 2026 · Issue 048.