Ling 2.0 scales open MoE language models from 16B to 1T parameters

The Ling Team posted Ling 2.0, a reasoning-oriented open language foundation that scales from tens of billions to one trillion parameters under a single Mixture-of-Experts (MoE) design, in a paper submitted to arXiv on 25 October 2025 and revised on 7 November. The series ships three non-thinking, or instruct, models: Ling-mini-2.0, Ling-flash-2.0 and Ling-1T, spanning 16B to 1T total parameters.

The organizing claim of the release is stated in the title: every activation boosts reasoning capability. The team says it built the series around high sparsity, cross-scale consistency and efficiency guided by empirical scaling laws, rather than treating each model size as a separate design problem. The payoff the paper reports is up to 7-fold active-compute efficiency compared with dense counterparts, meaning the models activate a small fraction of their total parameters per token.

Ling 2.0 combines changes across four layers of the stack. On architecture, it uses a high-sparsity MoE with MTP, or multi-token prediction, aimed at efficient reasoning. On data and pre-training, the team describes reasoning-oriented data and mid-training CoT activation, a stage that surfaces chain-of-thought behavior before final fine-tuning. On post-training, it applies reinforcement-based fine-tuning methods it names DFT and Evo-CoT. On infrastructure, the models were trained at full scale in FP8 precision using fine-grained heterogeneous pipelines.

The headline result concerns the largest model. At the trillion-parameter scale, Ling-1T establishes what the team calls a new Pareto frontier of reasoning accuracy versus computational efficiency. The paper's argument is that sparse activation, when aligned with reasoning objectives, can hold accuracy while cutting the compute needed to reach it. The abstract states this as a demonstration rather than a projection.

The paper lists the Ling Team as first author alongside more than 140 named contributors, with 42 additional authors not shown in the arXiv listing. That author count reflects the infrastructure and data work the abstract describes, which spans training-pipeline engineering as much as modeling.

The release is an arXiv preprint. The abstract does not name benchmark suites, report per-task scores, or give the token counts and compute budgets behind the scaling laws it cites, so the accuracy claims cannot be checked against specific evaluations from the abstract alone. The 7-fold efficiency figure is stated relative to dense counterparts without specifying which dense baseline or at what scale the comparison was made.

What remains open is how Ling-1T performs against other open trillion-scale and frontier reasoning models on standard evaluations, and whether the cross-scale consistency claim holds at the smaller 16B end of the series. The paper's v2 revision on 7 November indicates the work is still being updated. Weights, licensing terms and inference requirements for a 1T-parameter MoE are not described in the abstract.