

More posts
New MoE architecture. 2.4T total params, 71B active, native vision, 1M token context. Uses roughly 50% fewer tokens than last gen, per-task cost down over 60%, priced at $1.5/$4.5 per million. On paper it's the flagship. In practice, it whiffed: AA Intelligence Index 43, Terminal-4.0 48.5. Lost to kimi-k3.1-pro on both.

AMD's China dev platform gives out free daily LLM credits. Been using it for a while. I get through roughly 20M tokens of ds a day before hitting 429s. Pretty stable so far, and faster than opencode go. Might just be that hardly anyone uses it yet.

Kitesurf has no Chrome and no container. It's Rust compiled to WASM, running directly on Workers. It can parse HTML, run page JS, render screenshots and PDFs. For agent browser automation it uses 3-7x less CPU and memory than Chromium. It was never open source before, only usable on Cloudflare Workers. Now it is: Apache 2.0, cloudflare/kitesurf on GitHub.

Rumor is Nous is working with Google directly on this one. Meaning you'd get Hermes calling your Gemini subscription with no ban risk, and honestly, Google AI plans just keep stacking up value. Probably not empty talk. The Hermes repo already ships a hidden Solstice provider pointed at Google's per-user-quota endpoint, client secret hosted by Nous Portal.
