← Back to feed

Ant Group’s inclusionAI lab open-sourced Ling-2.6-1T on April 30 — a 1-trillion-parameter model targeting agent deployment with a hybrid architecture designed to reduce token waste rather than extend reasoning chains.

The model sits in the growing cluster of open-source frontier-tier offerings: below the ceiling of GPT-5.5 and Claude Opus 4.7, but competitive with GPT-5.4 on execution-focused benchmarks. Its differentiation is not raw capability — it is output efficiency in production agent workloads.

The Architecture

Ling-2.6-1T combines Multi-head Linear Attention (MLA) with Lightning Linear Attention, a design Ant calls a “Fast Thinking” mechanism. The goal is to suppress what the team describes as process redundancy: the tendency of large models to emit extended preambles and verbose chains before reaching an answer.

In practice, the claimed payoff is substantial. On complex reasoning tasks where comparable 1T-scale frontier models use 15,000–23,000 output tokens, Ling-2.6-1T is reported to produce equivalent-quality answers in approximately 5,000–6,000 tokens — a figure derived from the predecessor Ling-2.5-1T’s AIME 2026 performance and carried forward into the 2.6 generation.

That matters in production. Token costs scale with output length. A model that answers in 5K tokens where its peer uses 20K costs roughly 75% less per agent turn, assuming comparable input prompt sizes.

What It Is Built For

Ling-2.6-1T is explicitly targeted at enterprise agent deployment, not conversational AI. Training emphasis is on:

• Code modification and defect fixing
• Multi-tool invocation under complex constraint sets
• Long-context instruction following in noisy environments
• Composite task execution across enterprise workflows

Context window is 262,100 tokens, maximum output 32,800 tokens. The model supports native compatibility with mainstream agent frameworks and is described by Ant as a “core capability foundation for continuous enterprise deployment” — closer to an infrastructure component than a chat model.

Benchmarks

On Artificial Analysis’ Intelligence Index, Ling-2.6-1T scores in the upper tier of open-weight models. Independent comparisons on the AA platform place it competitively against Claude Opus 4.7 (Non-reasoning, High Effort) across code generation and long-context retrieval tasks, though the Anthropic model retains an advantage in multi-modal and broader general intelligence.

Ant reports top-level open-source performance across multiple execution-task evaluations for code generation, defect fixing, and instruction following under ambiguity. No independent SWE-bench Verified submission has been filed as of publication.

Context: The Ling Line

The Ling-2.6-1T is the flagship of Ant Group’s Ling (BaiLing) general-purpose model series. The companion Ring series handles deep reasoning tasks. Ling-2.6-Flash — the smaller, faster variant — attracted notice in April when it topped OpenRouter’s traffic charts as an unlabelled model before being identified. The 1T flagship is a different proposition: scale for execution depth, efficiency for cost-sensitive production, and full open weights for enterprise customisation.

Availability

• Hugging Face: inclusionAI/Ling-2.6-1T
• ModelScope: inclusionAI/Ling-2.6-1T
• OpenRouter: free API trial, extended by one week from April 30
• License: Ant Group InclusionAI open license