InclusionAI releases Ling-2.6-1T: 1 trillion parameter model free on OpenRouter with 262K context

InclusionAI has released Ling-2.6-1T, a 1 trillion parameter instruct model now available free on OpenRouter. The model features a 262,144 token context window and uses a "fast thinking" approach that the company claims reduces costs to roughly 25% of comparable models.

Technical specifications

• Parameters: 1 trillion
• Context window: 262,144 tokens
• Pricing: $0 per million input tokens, $0 per million output tokens (free tier on OpenRouter)
• Release date: April 23, 2026
• Model type: Instruct (instant) model

Performance claims

According to inclusionAI, Ling-2.6-1T achieves state-of-the-art results on AIME26 and SWE-bench Verified benchmarks. Specific scores were not disclosed in the release information. The company positions the model for "advanced coding, complex reasoning, and large-scale agent workflows."

The model's "fast thinking" architecture is designed to prioritize execution speed and efficiency over the extended reasoning approaches used by models like OpenAI's o1 series. InclusionAI claims this approach delivers performance comparable to top-tier models while operating at approximately 25% of the computational cost.

Availability

Ling-2.6-1T is currently available exclusively through OpenRouter's free tier. OpenRouter routes requests across multiple providers to maximize uptime and handle varying prompt sizes. The model uses OpenRouter's normalized API, compatible with OpenAI and Anthropic SDKs.

InclusionAI describes the model as suitable for "real-world agents that require fast execution and high efficiency at scale," positioning it for production deployments where inference costs are a primary concern.

What this means

Ling-2.6-1T represents a significant release in the trillion-parameter model space, particularly with its free availability through OpenRouter. The 262K context window places it among the longest-context models available. However, the lack of disclosed benchmark scores makes it difficult to verify performance claims against established models. The "fast thinking" approach appears to be a direct response to the high computational costs of reasoning models, targeting users who prioritize speed and cost over extended reasoning capabilities. If the efficiency claims hold, this could make large-scale agent deployments more economically viable.