Solar Mini 4: Upstage's 3B-active agent model, 512K context, $0.05/M
A 35B mixture-of-experts that bills like a tiny one — plus a zero-token Jev endpoint beside it, Aion 3.5 for storytelling, and Vercel's 16 TiB sandbox drives.
Copy markdown
A 3B-active MoE that costs like a nano model
Solar Mini 4 packs 35B parameters but fires only 3B per token, with a 512K context and 128K max output. It's $0.05/$0.20 per million through Oct 22 (list $0.10/$0.40, cached input $0.01), live now on Upstage Console with no waitlist — cheap enough to run the agent loops that would bankrupt you on a frontier model.
Update: a zero-token 'Jev' endpoint ships beside it
Alongside the model, Upstage reportedly added Solar Jev — a decision endpoint that returns a choice, score, or yes/no with a calibrated probability and bills zero output tokens. It's the same System-One pattern builders have been wiring in as probability gates, now with a hosted option: veto a risky shell command or route a request without paying to decode prose.
Aion 3.5 turns a model committee into one storyteller
AionLabs shipped Aion 3.5, a GLM-based system that fans each prompt across specialized models for tighter narrative structure, with a 262K context at $3/$6 per million. If you're building roleplay or interactive-fiction apps — its top users are Janitor AI and SillyTavern — it's a drop-in via OpenRouter.
Vercel Sandbox gets persistent drives
Vercel Sandbox now offers persistent drives up to 16 TiB in public beta, billed by usage. Your sandboxed agents can finally keep state — clones, caches, build artifacts — across runs instead of starting cold every time.