Product Hunt
Resurf took the top spot on Product Hunt this week with 295 votes — a Mac-native personal context library that captures and resurfaces what you've seen across apps, addressing the "where did I see that?" fragmentation that hits anyone juggling a dozen tools. It launched alongside Perplexity's Hybrid Compute (230 votes), which shipped September 1 and rolled out broadly this week: a Mac feature that splits agent tasks between the cloud and an on-device model (PPLX Qwen 3.8 27B on Apple Silicon), routing anything touching bloodwork, tax documents, or litigation steps to local hardware instead of the cloud. Perplexity also open-sourced the classifier that decides which is which.
The rest of the week leaned hard into Mac-native and privacy-first. Google shipped a Windows AI assistant of its own — Gemini for Windows, summoned with Alt+Space — signaling the OS-level assistant race is now a two-way fight, not just an Apple/Microsoft story. Developer tooling stayed busy too: Cortex and QApilot MCP for Android both launched around automating API documentation and in-code testing, and Youkti, an AI sales-intelligence tool that scores leads and prescribes next steps, pulled 332 votes — the highest of the week across any category.
Elsewhere
Cognition launched SWE-2 on September 10, its successor to SWE-1.7 and the company's closest model yet to frontier-tier coding performance. The headline: 50.0% on FrontierCode 1.1 Main, within one point of Claude Fable 5.1's 50.9% — while Cognition claims a 64% cost advantage at that score level, and reportedly a similar gap against GPT-6 Astra. It's a post-train of Moonshot AI's open-weight Kimi K3, which matters as much as the benchmark: the fastest way to a near-frontier coding model right now is fine-tuning someone else's open weights, not training from scratch. That's the second coding-model price shot fired this month, and it's putting real pressure on what Cursor, Windsurf, and the rest charge per token.
On GitHub, the trending list this week has a clear theme: local-first is winning mindshare over cloud-dependent. JustVugg/colibri picked up +2,173 stars in a single day — a zero-dependency, pure-C engine that runs frontier MoE models (Mixtral, DeepSeek-MoE class) by streaming experts from disk, meaning 10B+ parameter models can run on a laptop with no GPU. Alongside it, a fully local ElevenLabs-alternative voice synthesis project (VoiceStudio) gained similar traction supporting 646 languages. Two different domains, same bet: model owners betting that "runs entirely on your machine" beats "runs faster in someone's datacenter" for a growing slice of users.
Worth your time
If you only look at one thing this week, make it Cognition's SWE-2 — not because it's the best coding model out, but because it's proof that the fine-tune-an-open-model playbook can get within a point of frontier performance at less than half the cost. That's the number that should worry anyone pricing a coding agent on a per-token markup. Colibri is the sleeper: if disk-streamed MoE inference on commodity hardware holds up outside a demo, it changes the economics of who can run a serious model at all.