ling-2.6-1t

Ling-2.6-1T is a trillion-parameter flagship built for complex reasoning, coding, daily workflows, and tool-based tasks. Its architecture combines MLA with Linear Attention on a highly sparse mixture-of-experts backbone, an approach designed to reduce latency, memory use, and the cost of processing long contexts. The model is shipped in FP8 and is documented to fit on a single GB300 node with four-way tensor parallelism.

The model emphasizes practical agent performance through improved instruction following, multi-step planning, and execution-focused coding behavior. Post-training introduces “Fast Thinking” with Contextual Process Redundancy Suppression, which is intended to shorten reasoning traces and reduce token overhead while preserving answer quality. InclusionAI reports open-source SOTA results on AIME26, SWE-bench Verified, BFCL-V4, TAU2-Bench, and IFBench, making this model best suited to demanding engineering and workflow automation rather than lightweight general chat.

Requestyling-2.6-1tling

Open provider page Provider docs

Quick Info

Provider: Requesty
Model key: ling-2.6-1t
Release date: Apr 23, 2026
Last updated: Apr 23, 2026
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.30
Output token cost: $2.50

Limits

Output tokens: 262,144 tokens
Context window: 262,144 tokens

Transparent token rates

Compare ling-2.6-1t pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about ling-2.6-1t

Requesty

Coverage

Ant Group's Bailing Large Model team officially open-sourced the trillion-parameter flagship Ling-2.6-1T on April 30, 2026, positioning the model around instruction execution, tool adaptation, and long-context handling rather than raw parameter stacking. The model is now available on Hugging Face (huggingface.co/inclus Ling-2.6-1T uses an innovative hybrid architecture with an enhanced reward strategy that suppresses process redundancy, enabling a "fast thinking" mechanism that cuts token cost while preserving trillion-parameter intelligence quality. The model reaches top-tier open-source results on execution-oriented benchmarks incl

Requesty

CoverageBenchmark

BenchmarkList compiles third-party evaluation data for InclusionAI's Ling-2.6-1T, reporting a Tau2-Bench Telecom success rate of 89.8% (84th percentile), Terminal-Bench Hard at 31.1% (79th percentile), a GDPval-AA Elo of 1045 (73rd percentile), and a ClawProBench final score of 57.4 (53rd percentile), with leader-field The page situates Ling-2.6-1T against Fable 5.1, Claude Opus 5, Kimi K3, Qwen3.8-2.4T-A95B, Qwen3.8-Flash-Next, and GLM 5.3/Flash across Agentic (5 evals, 73rd percentile median) and other categories, drawing scores from official model cards, Artificial Analysis, and public leaderboards dated May–June 2026. The data pr

Requesty

Coverage

According to AIBase's April 30, 2026 report, Ant Group's BaiLing (Ling) team officially open-sourced Ling-2.6-1T as a trillion-parameter flagship model focused on "Fast-Thinking" efficiency rather than stacking parameters or ultra-long reasoning chains. The article describes an innovative hybrid MLA + LinearAttention a The same report says Ling-2.6-1T reached parity with GPT-5.4 in non-inference mode on overall intelligence and completed the full Artificial Analysis evaluation using only about 16M tokens, with output costs roughly one-quarter of comparable models. These are aggregator-sourced claims, not an Ant Group/Inclusion AI pri