Ling-2.6-1T

Overview

Ling-2.6-1T is the open-weight flagship of Ant Group's Ling line, released in late April 2026 by the company's inclusionAI (Bailing) team. Ant's own documentation states it plainly: "Ling-2.6-1T is the latest-generation flagship model in the Ling series."

A note on the date: Ant publishes no dated release announcement. The Hugging Face repository inclusionAI/Ling-2.6-1T was created on 2026-04-29 at 03:19:36 UTC, which is the earliest primary-source timestamp available. Secondary coverage dates the open-source release to April 30, 2026, and this page uses that date. The two are a day apart; treat the repo timestamp as the hard fact and April 30 as the reported one.

That distinction matters. Ling is a family, not a model. Ant unveiled the Ling AI model family in October 2025 alongside Ling-1T, its first trillion-parameter language model. The family has since split into three lines — Ling (token-efficient, non-thinking), Ring (explicit reasoning), and Ming (multimodal) — each with its own release cadence, and each with its own section in Ant's developer documentation. Ling-2.6-1T is the current head of the Ling line.

Ring-2.6-1T is a different model, not a rename of this one. Both repositories exist on Hugging Face, both are MIT-licensed, and both are trillion-parameter MoE models with 63B activated parameters — Ant's docs say Ring "uses the same MoE (Mixture of Experts) architecture as Ling." What separates them is what the model does with its tokens. The 2.6 technical report draws the line: "Ling-2.6 is optimized for instant response generation and high capability per output token, whereas Ring-2.6 is tailored for deeper reasoning and more advanced agentic workflows." Ring-2.6-1T landed on Hugging Face on 2026-05-14, two weeks after Ling-2.6-1T, and Ant's rate card lists both — Ling-2.6-1T as "Flagship," Ring-2.6-1T as "Flagship reasoning." Neither has superseded the other. If you are comparing third-party scores, check which line they belong to; the two are routinely confused.

The model is a sparse Mixture-of-Experts transformer: approximately 1T total parameters with 63B activated per token. Two design choices define it. First, "a hybrid architecture combining MLA and Linear Attention," which Ant says "dramatically reduce[s] latency and VRAM footprint for long contexts." Second, and more unusual, a post-training objective aimed squarely at output cost.

The Fast Thinking bet

Most frontier models buy accuracy with thinking tokens. Ling-2.6-1T tries to buy it back. From the model card:

"We introduce a Contextual Process Redundancy Suppression reward strategy during post-training. This reduces reliance on verbose chains-of-thought (CoT), utilizing a 'fast thinking' mechanism to reach answers directly and compress output costs while maintaining top-tier intelligence."

Ant's docs describe the result as "fast thinking replaces lengthy reasoning chains, preserving 1T parameter level capability at lower token cost." This is a bet that in production, tokens-per-solved-task matters more than accuracy at any token budget — a reasonable position for a company that runs Alipay-scale inference.

Capabilities

• Fast Thinking: Trained to reach answers directly rather than through long chain-of-thought traces, compressing output-token spend.
• Execution-heavy agentic work: Ant claims "open-source SOTA on execution-heavy benchmarks, including AIME26, SWE-bench Verified, BFCL-V4, TAU2-Bench, and IFBench."
• Software engineering: 72.20 on SWE-bench Verified, under the openhands scaffold.
• Million-token native context: Supported natively by the hybrid MLA / Linear Attention stack, though the hosted API currently exposes 256K.
• Agent framework compatibility: Ant's docs highlight compatibility with mainstream agent frameworks and strong tool adaptation.
• Long-context efficiency: Linear Attention layers cut VRAM footprint, which is what makes the long context practical rather than nominal.

Technical Specifications

• Hugging Face repository: inclusionAI/Ling-2.6-1T
• Total parameters: approximately 1T
• Activated parameters: 63B per token
• Architecture: sparse MoE with a hybrid MLA + Linear Attention stack
• Post-training: Contextual Process Redundancy Suppression reward strategy ("Fast Thinking")
• Context window: 1M tokens native; 256K exposed via Ant's API; config.json sets 262,144 positions
• License: MIT
• Serving: vLLM and SGLang

Ant does not publish a knowledge cutoff or a training-token count for Ling-2.6-1T, so this page does not state one.

The rest of the family

Ling-2.6-flash targets customer service, content generation, real-time translation, and semantic understanding; Ant claims "aggregate benchmark performance comparable to or exceeding 40B-class Dense models" with superior throughput. Ring-2.6-1T is the explicit-reasoning counterpart: the same trillion-parameter MoE architecture with 63B activated, published to Hugging Face on 2026-05-14, but with a 256K context window (its config.json sets 131,072 positions, extended to 256K via YaRN) and an adjustable Reasoning Effort control — high for agent workflows, xhigh for hard mathematics and analysis — in place of Ling's Fast Thinking.

The previous generation — Ling-2.5-1T and Ring-2.5-1T — was released on February 15, 2026.

Use Cases

• Cost-sensitive agents: The whole design premise. If your agent loop pays per output token, a model trained to skip redundant reasoning changes the unit economics.
• Multi-step task execution: Ant's docs list four use cases for Ling-2.6-1T — "Multi-step tasks and agent tasks"; "Code completion, project delivery, and bug fixing"; "Data visualization projects such as slides and reports"; and "Long-context knowledge management and automated workflows."
• Software engineering: 72.20 SWE-bench Verified (openhands scaffold) with an MIT license and a 256K working window.
• Tool-using agents: Compatibility with mainstream agent frameworks, with BFCL-V4 and TAU2-Bench cited among the claimed open-source SOTA results.
• Long-context knowledge management: The native 1M window is available to self-hosters willing to extend beyond the API's 256K.
• High-throughput serving — via Ling-2.6-flash: At 7.4B active parameters, the flash variant is the right tool for classification, translation, and customer-service volume.

Performance / Benchmarks

Reported in the benchmark chart on Ant's Ling-2.6-1T model card.

Two things about this table are easy to miss. First, SWE-bench Verified is run under the openhands scaffold, which Ant labels explicitly in the chart; SWE-bench scores are scaffold-dependent and are not comparable across harnesses. Second, the entire comparison set is non-reasoning: Ant benchmarks Ling-2.6-1T against GPT-5.4, GLM-5, DeepSeek-V3.2, and Kimi-K2.5, each labelled "(Non Reasoning)" in the chart legend. That is a defensible choice for a model whose selling point is answering without long chains of thought, but it means these results say nothing about how Ling-2.6-1T compares to the same models in reasoning mode.

Third-party: the model card cites an Artificial Analysis Intelligence Index score of 34, produced with roughly 16M output tokens. Ant does not name the index version. That score dates from the April 2026 release, when Intelligence Index v4.0 was current; Artificial Analysis shipped v4.1 in June 2026, replacing several evaluations and reweighting the categories, so the 34 does not carry across.

Against Artificial Analysis's current published numbers (Intelligence Index v4.1, as of 2026-07-08):

If you arrived at this page from an Artificial Analysis score of ~30.6, that score is Ring-2.6-1T's, not Ling-2.6-1T's. The two are separate models. The gap is the point of the split rather than a defect: Ring reaches 31 by spending roughly 100M output tokens on the index, while Ling is built to answer without them — which is also what Ant's 16M-token figure is advertising.

Vendor-reported figures should be treated as an upper bound. Validate on your own evaluation set.

Limitations

• Benchmarks are published against a non-reasoning field only: The eight published scores compare Ling-2.6-1T with GPT-5.4, GLM-5, DeepSeek-V3.2, and Kimi-K2.5, all in non-reasoning configurations. There is no vendor-published comparison against reasoning-mode models.
• API context is 256K, not 1M: The native window is 1M, but you only get it by self-hosting and configuring for it. config.json ships with 262,144 positions.
• Headline pricing is promotional: The ¥0.45 / ¥1.80 rate is 1/10 of list and expires 2026-07-31. Budget against the ¥4.50 / ¥18.00 list price unless Ant extends the promotion.
• Fast Thinking is a trade: Suppressing chain-of-thought reduces token spend. On problems that genuinely need long deliberation, the Ring line exists for a reason.
• Self-hosting scale: A trillion-parameter MoE requires a substantial multi-GPU cluster even at 63B active parameters. Ling-2.6-flash is the practical self-hosting target for most teams.
• No published knowledge cutoff: Ground time-sensitive queries with retrieval or tool use.
• Weights, not full openness: MIT covers the weights. Training data and pipeline are not published.

Pricing & Access

Self-hosting is free. Weights are MIT-licensed and published on Hugging Face and ModelScope. Commercial use, modification, and redistribution are permitted.

Hosted access is sold through Ant's developer platform, which publishes a rate card. Ant states: "All prices below are in CNY."

The promotional rates are not permanent. Ant's page states that during the promotional period, "valid until 2026-07-31," "all model prices are 1/10 of the original price. Original prices resume after the promotion ends." Each account also receives 500,000 free tokens per day, shared across input and output, resetting at 2:00 AM UTC+8.

These are yuan prices, quoted in CNY by the vendor. No USD conversion is implied.

Ant separately publishes USD rates for two third-party platforms, noting "All prices below are in USD." On ZenMux, Ling-2.6-1T is $0.30 input / $2.50 output / $0.06 cache read per million tokens. On OpenRouter, the same list prices apply, discounted during the promotion to $0.075 / $0.625 / $0.015 — Ant notes that on OpenRouter, "Ling-2.6-1T and Ring-2.6-1T are offered at 25% of the original price, and Ling-2.6-flash at 10% of the original price." Ant adds that OpenRouter rates are "subject to the official OpenRouter pricing."

Serving: vLLM and SGLang are the standard paths for the Ling and Ring weights.

Ecosystem & Tools

• Ling-2.6-1T on Hugging Face — flagship weights and model card
• Ling-2.6-flash on Hugging Face — 104B / 7.4B active efficiency model
• Ring-2.6-1T on Hugging Face — the explicit-reasoning counterpart
• inclusionAI on Hugging Face — the full Ling, Ring, and Ming families
• ModelScope — mirror for faster downloads in China
• Ant Ling developer docs — model descriptions and API guidance
• Ant Ling pricing — the published CNY rate card, plus Ant's USD rates for OpenRouter and ZenMux
• vLLM / SGLang — high-performance local serving

Community & Resources

• Ant Ling model documentation
• Ant Group Unveils Ling AI Model Family and Launches Trillion-Parameter Language Model Ling-1T — the October 2025 family announcement
• Ant Group Releases Ling-2.5-1T and Ring-2.5-1T — the February 2026 previous generation
• Ling-2.6-1T model card
• Compare with DeepSeek V4, Kimi K2.6, GLM-5.2, and Tencent Hy3