MDL-4066EST.2026 · IDX.708
Language modelIn production

Ring-2.6-1T

InclusionAI (Ant Group) · China · 2026

Ant Group's trillion-parameter open-weight reasoning flagship, shipped in FP8, with two selectable reasoning depths and no third-party hosting.

wujec.ai score

/10

Community score

no votes yet
Sign in to rate

Ring-2.6-1T is the reasoning flagship of InclusionAI, the open-source arm of Ant Group, and the largest model that publisher has released. The weights went up on Hugging Face on 14 May 2026, the same day the repository was created, in 175 shards; the published safetensors index counts 1.026 trillion parameters. InclusionAI does not state how many of them run per token, so this catalogue does not quote an activated figure. The model is not a fresh training run. The technical report published a month after the weights, on 13 June 2026, describes Ring-2.6 as an upgrade of the older Ling-2.0 base through architectural migration pre-training followed by large-scale post-training. The architecture is a hybrid: Lightning Attention interleaved with multi-head latent attention across 80 layers, 256 routed experts with 8 activated and one shared, four dense layers at the front, hidden size 8192 and a 157,184-token vocabulary. Context is 131,072 tokens natively and stretches to 262,144 with YaRN. The weights are distributed in FP8 rather than converted after the fact. The distinctive feature for a user is a reasoning effort switch with two settings. The high setting is meant for everyday agent workflows and keeps reasoning token overhead down; xhigh opens the model up for mathematics, research and multi-path analysis. The maker publishes separate benchmark numbers for each: on high, 87.60 on PinchBench, 63.82 on ClawEval and 95.32 on the telecom scenario of Tau2-Bench; on xhigh, 66.18 on ARC-AGI-V2, 95.83 on AIME 26 and 88.27 on GPQA Diamond. All of these are the maker's own measurements. The licence position needs stating plainly. The repository metadata carries an MIT label and the model card says the code repository is licensed under MIT, linking to a file in a separate GitHub project. Among the 190 files in the weights repository there is no licence document at all. The same arrangement appears in this publisher's other releases, so it is a house habit rather than a one-off omission. One practical consequence shows up in the download counts. Ring-2.6-1T was fetched 493 times in the 30 days to 29 August 2026; its predecessor Ring-2.5-1T, a model of almost identical size, was fetched 9,366 times in the same window. Neither is resold by any provider on OpenRouter, and the deployment instructions still point at a custom branch of SGLang rather than an official release. InclusionAI runs the model itself for free at ling.tbox.cn.

#open weights#MoE#reasoning#agentic#trillion parameters#FP8#256K context
Official website

Videos

No videos yet.

Reviews

No reviews yet. Be the first!

Sign in to write a review