Update: open Ling 3.0 lands free on Vercel AI Gateway, adds Tiny
Two weeks after debut, Ant's 124B MoE is free with no markup till mid-August and gains a free Tiny variant — plus fresh open drops from NVIDIA and Microsoft.
Copy markdown
The delta: free on Vercel AI Gateway till mid-August
Ling-3.0-flash — the open 124B MoE we covered at launch — is now on Vercel AI Gateway with no markup, BYOK, high rate limits, and automatic failover, free for roughly three weeks from Aug 6. Point Cursor, Claude Code, or any OpenAI-compatible client at it and A/B it against your current model for nothing.
A free Tiny sibling joins the family
Ling 3.0 Tiny dropped alongside the AI Gateway push, free on OpenRouter with open weights on Hugging Face. Route drafts and high-volume calls to Tiny and save Flash for the hard reasoning — a two-tier setup that costs almost nothing.
The refresher: MIT weights that actually code
For anyone who skipped launch: 124B total but only 5.1B active per token, MIT-licensed, 256K context, scoring 56.6 on SWE-Bench Pro and 93.2 on AIME 2026. It matches Ant Group's older 1T-param model at a fraction of the compute, so self-hosting is realistic.
NVIDIA opens a full-duplex voice model
NemotronLabs VoiceChat 11B listens and speaks at once and can fire tool calls mid-conversation, with ~450ms turn-taking. The catch: an OpenMDW research-only license and an 80GB-VRAM floor, so it's a prototype base to study, not a shippable product yet.
Microsoft's Mage-VL treats video like a codec
Mage-VL is a 5B Apache-2.0 vision-language model that splits video into anchor and predicted frames, cutting visual tokens ~75% for up to 3.5x faster inference. Self-host it for real-time video commentary or streaming analysis without a per-frame token bill.
Vercel's Chat SDK can now pause for a human
A single requestApproval call posts an approval card and suspends an agent workflow until you decide, persisting through deploys and restarts. It's the human-in-the-loop primitive agent builders keep hand-rolling.