NaiveAI releases Naive-N0.5-Flash: 309B open MoE, MIT, 1M context
A frontier-scale coding model you can run and fine-tune for free — plus Perplexity's reusable Agent API and Manus 2.0's Cascade harness and Cue agents.
Copy markdown
309B open weights under MIT, tuned for code
NaiveAI open-weighted Naive-N0.5-Flash: a 309B mixture-of-experts with ~15.5B active params, a hybrid SWA–DSA attention design (no full-attention layers) and native 1M-token context, built on MiMo-V2.5 for coding and AI R&D. It's on Hugging Face under MIT, so you can serve, quantize, or fine-tune a frontier-scale coder with no API bill or usage limits.
Perplexity Agent API: define an agent once, reuse everywhere
Perplexity's Agent API now saves an agent's model, instructions, tools, Skills, and managed connectors under one versioned Profile ID you can call from any app. Managed connectors ship for GitHub, Slack, Google Drive, Linear, Notion, and Datadog — so a shared agent stops meaning copy-pasted prompts and hand-rolled OAuth.
Manus 2.0 ships the Cascade harness and Cue agents
Manus 2.0 (Sept 28) introduced Cascade, an agent harness that loads specialized capabilities only when a task needs them — Manus reports one internal run at 23.2% fewer tokens, 28.2% faster, and 32% cheaper than its old stack. Companion app Cue gives each agent its own email, phone number, and wallet, and is in free early access with invite code MEETCUE.