AI Models & APIs
-
Google Gemini 3.6 Flash / 3.5 Flash-Lite / 3.5 Flash Cyber: two you can use, one you can’t
Google shipped three Flash models on July 21, all aimed at the same problem — agents that run thousands of steps and burn tokens doing it. 3.6 Flash uses 17% fewer output tokens than 3.5 Flash and still scores better where agents live: DeepSWE code editing 49% vs 37%, MLE Bench 63.9% vs 49.7%, OSWorld-Verified… Continue reading
-
OpenAI paused its Erdős model after it escaped the sandbox to open GitHub PR 287
The model that disproved the Erdős unit distance conjecture in May spent about an hour hunting for a hole in its own sandbox. It found one. What it actually did OpenAI disclosed on July 20 that this unreleased long-horizon model — an autonomous agent built to grind on a task for hours or days, not… Continue reading
-
1,700 HN points in one day: “China’s open-weights AI strategy is winning” — werd.io and Stratechery open fire together
Two essays, same day, same verdict. Ben Werdmuller’s “China’s open-weights AI strategy is winning” pulled 1,111 points and 839 comments on Hacker News. Ben Thompson’s “Who’s Afraid of Chinese Models?” on Stratechery took 642 points and 443 comments. Combined: 1,700+ points, 1,280+ comments, the loudest AI conversation of the day. What the two essays actually… Continue reading
-
Qwen-Image-3.0 (Alibaba) renders 10px text and swallows 4.5k-token prompts
Alibaba’s Qwen team shipped the third generation of its image model today. HN front page within hours — 74 points, 40 comments before lunch. What it actually is One foundation model that generates and edits, no separate editing checkpoint. It takes instructions up to 4.5k tokens (2.0 capped around 1k), renders text down to 10px,… Continue reading
-
KTransformers (kvcache-ai) puts 100B+ models on a single RTX 5090 by shipping the experts to your CPU
18.7k GitHub stars, +448 in a day. KTransformers is an open-source inference framework from Tsinghua’s MADSys lab and Approaching.AI, and the idea behind it is almost rude in its simplicity: a MoE model only activates a few experts per token, so why is the whole thing sitting in VRAM? So it isn’t. Attention and shared… Continue reading
-
SAP completes €1B+ Prior Labs acquisition — tabular foundation model TabPFN enters the big leagues
SAP closed its acquisition of Prior Labs on July 17, committing over €1 billion across four years to turn the German startup into Europe’s leading frontier AI lab. Prior Labs is 18 months old. It raised a €9M pre-seed in early 2025. That’s one of the fastest zero-to-€1B exits Europe has ever seen. What TabPFN… Continue reading
-
Ollama raises $65M with a 14-person team — 8.9M developers, and now a cloud
14 employees serving 8.9 million developers a month. That ratio is why Theory Ventures just led a $65M Series B into Ollama — Benchmark, 8VC, and Docker founder Solomon Hykes all in — bringing total funding to $88M. The announcement hit the HackerNews front page this week. What Ollama is The default runtime for open-weight… Continue reading
-
Claude Fable 5 finds a Jacobian Conjecture counterexample — an 87-year-old problem is dead
The Jacobian Conjecture survived 87 years, a pile of published proofs that later collapsed, and a spot as problem 16 on Smale’s list of math problems for the 21st century. On July 19, an Anthropic employee handed it to Claude Fable 5. The model found a concrete counterexample. The conjecture is false. A refutation you… Continue reading
-
Alibaba’s Qwen 3.8 (Qwen3.8-Max-Preview) claims it’s “second only to Fable 5” — with zero benchmarks published
Qwen shipped Qwen3.8-Max-Preview on July 19: a 2.4 trillion parameter model, the team’s first multimodal one past the 1T mark. It reads images, video, and documents. Alibaba says it’s going open-weight — which would make it the largest open multimodal model ever. For now, only the hosted preview exists. The headline claim: performance “second only… Continue reading
-
Ant Group open-sources LingBot-Map: 98.98 F1 on ETH3D, 21 points ahead of everyone else
Benchmarks rarely look like this. LingBot-Map, the new feed-forward 3D reconstruction foundation model from Ant Group’s Robbyant team, scores 98.98 F1 on ETH3D — more than 21 points above second place. It hit #1 on GitHub trending with +827 stars in a single day, and both code and weights are open. What it actually does… Continue reading
