Ling 3.0 Flash: an MIT coding model, free on Vercel through Oct 4
A 124B/5.1B-active MoE hitting 56.6% on SWE-Bench Pro that drops into Claude Code, free on Vercel now — plus Gemini 3.8 Flash cuts long-context pricing.
Copy markdown
124B open weights, 5.1B active
InclusionAI's Mixture-of-Experts fires just 5.1B of 124B parameters per token and ships under an MIT license on Hugging Face — commercial use, fine-tunes, and self-hosting with no strings attached.
Built for coding agents
It posts 56.6% on SWE-Bench Pro and 72.4% on SWE-Bench Multilingual with reasoning on by default, and it's tuned to drop straight into Claude Code and Qwen Code as a low-cost backend.
Free through Oct 4, then pennies
Vercel added it to the AI Gateway free until October 4. After that, OpenRouter routes it around $0.02 in / $0.06 out per million tokens — cheap enough to leave an agent looping all weekend.
Why it's cheap to run
A hybrid-linear design (Kimi Delta plus latent attention) with built-in hierarchical caching cuts time-to-first-token 60-80% on long inputs, and the 256K window swallows a big repo in one pass.
Elsewhere: another cut-rate long-context pick
Gemini 3.8 Flash is now GA at $0.38 in / $1.88 out per million with a 1M-token window and live on the Vercel AI Gateway — a second low-cost option for agent and RAG workloads this week.