Ant Group's AI lab, Inclusion AI, just dropped the weights for its flagship Ling-2.6-1T model. It's a monster: 1 trillion total parameters, but only 63 billion are activated per inference thanks to a Mixture of Experts architecture. Context window: 256K tokens. License: MIT. You can download it right now.
That's a lot of numbers. But the real story is what this model is built to do: think fast and shut up. Yesterday, the lab released Ling-2.6-flash, a lightweight variant with 104 billion parameters and 7.4 billion active, designed for speed. The 1T version is for the tough stuff — and Inclusion AI claims it's the best open-source non-reasoning model on agent benchmarks.
Architecture-wise, it's the same hybrid attention as the flash version — a mix of MLA (multi-head latent attention) and Lightning Linear layers. But the 1T adds a "fast thinking" training trick: a reward that penalizes redundant chain-of-thought rambling. The idea is to squeeze reasoning into fewer tokens without dumbing down the output.
How does it actually perform? On SWE-bench Verified, it hit 72.2 percent (the flash version got 61.2). Inclusion AI says it's the top open-source non-reasoning model on execution benchmarks like AIME 2026, BFCL-V4, TAU2-Bench, and IFBench. For comparison, reasoning models like DeepSeek V4 Pro are already above 80 percent on SWE-bench. So Ling-2.6-1T is competitive, but not at the top of the chain-of-thought pile.
Still, it packs a punch. Artificial Analysis gives it a composite intelligence score of 34, chewing through about 16 million output tokens. The model plays nice with agent frameworks like Claude Code, OpenClaw, and OpenCode, and there's already a free API on OpenRouter. Minimum deployment: 8 GPUs, with support for SGLang and vLLM.