Ling-1T Now Live on SiliconFlow: A Trillion-Scale Leap in Efficient Reasoning
Oct 16, 2025
TL;DR: Ling-1T, the first flagship non-thinking model in the Ling 2.0 series, is now live on SiliconFlow. Built for efficient reasoning at trillion scale, it features 1 trillion total parameters with approximately 50 billion active per token and Evo-CoT training for deeper, faster reasoning. Delivering state-of-the-art performance across math, code, and front-end tasks, Ling-1T redefines the balance between accuracy, speed, and cost-efficiency.
Start building today with SiliconFlow's production-ready API, fully OpenAI/Anthropic-compatible and ready to integrate into your existing workflows!
Today, we're launching Ling-1T on SiliconFlow — the first flagship non-thinking model in the Ling 2.0 series. Built with 1 trillion parameters and ~50 billion active per token, it's optimized for efficient reasoning across complex tasks. It also excels at transforming complex natural language descriptions into executable code, solving competition-level math problems, and generating front-end interfaces that balance aesthetics and functionality. More importantly, it achieves higher accuracy with shorter reasoning chains, which translates to faster responses, lower costs, and more reliable outputs.
With SiliconFlow's Ling-1T API, you can expect:
Competitive Pricing: Ling-1T $0.57/M tokens (input) and $2.28/M tokens (output).
Extended Context Window: 131K context window allows you to process longer documents and maintain context across complex tasks.
Efficient MoE Architecture: 1 trillion total parameters with ~50 billion active parameters per token, delivering powerful reasoning without the computational overhead of fully dense models.
Evo-CoT Optimization: Pre-trained on 20 trillion+ high-quality, reasoning-dense tokens with evolutionary chain-of-thought (Evo-CoT) processes that enhance multi-step problem-solving.
Key Features & Benchmark Performance
Reasoning is the cornerstone of intelligence. Ling-1T pushes this further by combining efficient reasoning with creative generation, excelling across two key dimensions:
Aesthetic Understanding & Front-End Generation
Turn ideas into elegant front-end code: Ling-1T interprets UI intent and translates natural language into clean, functional interfaces.
Code that looks as good as it runs: Its hybrid Syntax–Function–Aesthetics reward system ensures every line of code balances correctness, usability, and visual appeal.
Ranked #1 on ArtifactsBench: Ling-1T excels in front-end generation. Scroll down to see demos created via the SiliconFlow API.
Emergent Intelligence at Trillion Scale
Scales with intelligence: Reveals emergent reasoning and transfer abilities at trillion scale, achieving ~70% tool-call accuracy on BFCL V3 with minimal tuning.
Understands your intent: Accurately interprets complex natural-language instructions, aligning intent with reasoning to generate reliable, goal-oriented responses.
Bridges logic and design: Turns abstract ideas or reasoning steps into functional visual components, helping you prototype faster with cleaner logic.
Builds across platforms: Generates responsive, cross-platform front-end code that's ready to deploy with minimal adjustments.
Writes with style and context: Produces marketing copy and multilingual text that align with your tone, audience, and brand identity.
To ensure a fair and comprehensive evaluation, Ling-1T was also benchmarked against both open-source and closed-source flagship models — including DeepSeek-V3.1-Terminus, Kimi-K2-Instruct-0905, GPT-5-main, and Gemini-2.5-Pro.
Across code generation, software development, competition-level mathematics, professional math, and logical reasoning, Ling-1T consistently delivers superior complex reasoning performance and demonstrates a clear overall advantage over these leading models.
On the AIME 25 benchmark, Ling-1T extends the Pareto frontier of reasoning accuracy vs. reasoning length, achieving higher accuracy with fewer reasoning steps.
For developers working on complex reasoning, mathematical analysis, or multi-step problems, this means faster results with higher precision.
Model Architecture & Training
Key architectural innovations:
Built on Ling 2.0 architecture and guided by the Ling Scaling Law, Ling-1T is designed from the ground up for trillion-scale efficiency, ensuring stable reasoning performance at any scale:
1T total / 50B active parameters with a 1/32 MoE activation ratio
MTP layers for enhanced compositional reasoning
Aux-loss-free, sigmoid-scoring expert routing with zero-mean updates
QK Normalization for fully stable convergence
Training Efficiency:
Ling-1T leveraged FP8 mixed-precision training, making it the largest known foundation model trained this way. Combined with heterogeneous pipeline optimization and the WSM learning rate scheduler, the training achieved over 40% end-to-end acceleration while maintaining stability at trillion-scale.
Pre-Training:
Pre-trained on 20 trillion+ high-quality tokens, with over 40% dedicated to reasoning-dense data. This foundation gives Ling-1T natural strength in logic, multi-step problem-solving, and complex analysis.
To further enhance these capabilities, curated chain-of-thought data was integrated during mid-training, improving reasoning stability and generalization.
Post-Training:
Ling-1T continues refining its reasoning ability through advanced optimization:
Evo-CoT (Evolutionary Chain-of-Thought): Progressively enhances reasoning precision and efficiency, enabling more logical thinking with fewer computation steps.
LPO (Linguistic-Unit Policy Optimization): Optimizes learning at the sentence level, aligning rewards with natural language meaning rather than token sequences.
Get Started Immediately
1. Explore: Try Ling-1T in the SiliconFlow Playground.
2. Integrate: Use our OpenAI-compatible API. Explore the full API specifications in the SiliconFlow API documentation.
Start building with Ling-1T today through SiliconFlow's production-ready API — delivering trillion-scale reasoning with efficiency and reliability.
Join our Discord community now →
