™All Models & Products
-
Anthropic: Claude Mythos 5 Autonomously Designs Quantum Circuits, One Design's Gate Count Cut to 1/17
Why it mattersThe model designed new circuits surpassing the best published ones, beating them on all eight tasks.
-
TypeSafe's Jev Becomes Vercel's Fastest-Adopted Model Ever, Reaching 35% of Paid Teams in Two Weeks
Why it mattersJev proved in two weeks that decision models that “choose without writing” have real scaled demand, while open-source models took over 70% of gateway token volume — the inference market is shifting from “bigger is better” to “good enough and cheap.”
-
Microsoft's MAI-Base-2 Completes Pre-Training: 2T-Parameter All-MoE Architecture, Trained 38 Days on 16,000 GB200s
Why it mattersMicrosoft published the full training ledger for a trillion-scale in-house base model for the first time, turning "training efficiency" into a new competitive dimension; the 2T-parameter in-house route means Microsoft no longer relies solely on OpenAI for frontier models.
-
Elorian Launches First Visual Reasoning Model Glimpse: Reasoning With Images
Why it mattersThe first model built around looking back at images mid-reasoning; a small team trained it on a few hundred GPUs to rival GPT-6 Astra on visual reasoning — a new route that doesn't rely on scale.
-
Claude Cracks 25-Year Cryptography Mystery in Half an Hour: NIST Curve Seeds Came From a 1997 Clock
Why it mattersThe first public case of AI autonomously making a cryptography-history-level discovery — a reproducible seed origin removes the ground for 25 years of "NSA backdoor" suspicion, and it took under half an hour.
-
DeepSeek Releases V4.1-Flash-1017: MIT-Licensed Weights, In-House Search API
Why it mattersOne release covers both an MIT-licensed flagship snapshot and an in-house search API; server-side Search/Fetch calls inside DeepSeek Harness, Claude Code and Codex are free until November 1.
-
DeepSeek V4.1-Flash-1017 Released: Weights Reportedly Open Under MIT License, In-House Search API Launches
Why it mattersOne release covers both model weights and an agent search toolchain, with the search API priced at $0.75 per thousand calls.
-
Reflection Releases Beam's Open Weights: Apache 2.0 License, 501B-Parameter MoE
Why it mattersEleven days after Beam's debut, Reflection delivered on its open-weights promise; a 501B-class release under Apache 2.0 is the US camp's latest answer to Chinese labs' lead on open-weight leaderboards.
-
Kling 4.0 Launches Globally: Native 30-Second Video Generation with 15 Reference Assets
Why it mattersKling packs native 30-second takes, keyframe storyboarding and 15 reference assets into one release — and lands in Adobe Firefly — moving AI video from short clips toward professional film production workflows.
-
Jailbroken PS5 Becomes a Local AI Server: Qwen3.8-27B Hits ~20 tok/s, Small Models Top 100
Why it mattersThe first time a game console has run a 27B model at usable speed through open-source porting alone — a 16GB, 448GB/s-bandwidth PS5 now sits in desktop inference territory, and the community just rewrote the hardware bar for local AI.
