Ling 1T: Open-Source Trillion-Parameter Intelligence
A production-grade foundation model from Ant Group's Bailing team for advanced reasoning workloads.
Mixture-of-Experts architecture with 1/32 routing · 50B active parameters per token · FP8 training for rapid convergence · 32K–128K context via YaRN
Deploy Ling 1T under the MIT license to power quantitative research, enterprise analytics, multilingual assistants, and secure agent ecosystems.
Or launch a specialized Ling 1T assistant
Explore curated Ling 1T prompts
Experience Ling 1T in Practice
Hands-on workflows that highlight Ling 1T's reasoning depth
Review curated walkthroughs that show how Ling 1T handles math proofs, complex codebases, and governance-ready reporting with traceable logic.
Benchmark Explorer
Inspect Ling 1T's performance on AIME 2025, MMLU-Pro, OlympiadBench, and more with annotated reasoning traces.
Training Methodology
Understand how FP8 mixed-precision, WSM scheduling, and Evo-CoT post-training deliver consistent gains.
Deployment Playbooks
Follow vLLM and SGLang setup guides, hardware sizing, and cost models for on-premise or managed inference.
Architecture Highlights
What sets Ling 1T apart from conventional large models
Engineered as a sparse Mixture-of-Experts system, Ling 1T balances trillion-parameter capacity with practical latency and enterprise governance.
Sparse MoE efficiency
Activates ~50B parameters per token with sigmoid expert gating to maintain speed without sacrificing depth.
- 1/32 expert routing keeps inference efficient without sacrificing comprehension.
- Only ~50B parameters activate per token while the full MoE remains available.
- Maintains latency comparable to 50B dense models across production workloads.
FP8 mixed precision
15% faster end-to-end training versus BF16 with negligible loss drift across 1T tokens.
- FP8 stack delivers ~15% faster training vs BF16 with negligible loss drift.
- YaRN scaling extends context from 32K to 128K without destabilising gradients.
- 1F1B pipeline maximises GPU utilisation while keeping memory predictable.
Extended context
32K token default context extends to 128K with YaRN so hierarchical documents remain intact.
- QK normalisation and fused kernels stabilise trillion-parameter optimisation.
- Checkpoint merging plus WSM scheduling smooths long-run convergence.
- Integrated telemetry surfaces routing and safety signals for auditing.
Stable optimization stack
WSM scheduling, checkpoint merging, and fused kernels maximize GPU utilization for large-scale runs.
- Sentence-level LPO alignment keeps reasoning grounded and transparent.
- Supports tool plans, long-context recall, and multilingual instruction following.
- MIT licence lets teams extend or self-host with full commercial freedom.
Deliver With Ling 1T
Translate trillion-parameter intelligence into reliable production behavior
Combine Ling 1T's reasoning stack with your data, tools, and guardrails to ship trustworthy AI assistants and copilots.
Engineering choices that matter
Ling 1T keeps inference efficient while preserving explainability through Evo-CoT and sentence-level LPO alignment.
How Ling 1T handles each request
- 1Ingest instructions, documents, and retrieval context up to 128K tokens without truncation.
- 2Route tokens across 32 experts so ~50B parameters engage per step with stable QK normalization.
- 3Return structured reasoning, tool calls, and citations optimized via Evo-CoT and LPO.
Production playbooks
Risk & Finance Briefings
Condense market, portfolio, and policy updates into investor-ready insights with transparent assumptions.
Review risk sampleTechnical Architecture Reviews
Transform requirements into architecture memos, dependency maps, and remediation plans.
Draft an architecture noteMultilingual Policy Support
Generate localized training, compliance FAQs, and customer communications with consistent terminology.
Localize guidanceWhere Ling 1T Excels
Operational scenarios proven by evaluation and production rollouts
Pair Ling 1T's trillion-parameter capacity with efficient 50B active routing to deliver disciplined reasoning, precise generation, and resilient tool use.
Quantitative Research
Reach 70.42% on AIME 2025 with concise derivations and reliable numeric reasoning across competition-grade problems.
Run a math analysisSoftware Engineering Automation
Top LiveCodeBench performance enables Ling 1T to draft, refactor, and review production services with clear rationales.
Generate review-ready codeEnterprise Knowledge Analysis
128K token context with YaRN unlocks long-form document parsing, synthesis, and regulatory reporting in one pass.
Summarize complex materialFinancial Risk Intelligence
In production at Ant Group, Ling 1T combines market data interpretation with scenario planning for high-frequency decision support.
Draft an investor briefingMultilingual Compliance
Ling 1T scores 92.19 on C-Eval, enabling policy reviews, training content, and localization across regulated markets.
Assess knowledge coverageAgent & Tool Orchestration
With ~70% accuracy on BFCL V3, Ling 1T coordinates tool calls, APIs, and workflows without heavy instruction fine-tuning.
Compose an agent planHow teams ship with Ling 1T
From prompt to production in three simple steps
Create an account, add funds, and use Ling 1T through hosted chat or compatible APIs with prepaid metering.
Connect & configure
Generate an API key or start in the hosted chat, then configure each request for your workload.
- Generate and revoke API keys from the dashboard.
- Choose personas and decoding controls in the hosted chat.
- Use OpenAI-compatible chat completions or Claude-compatible messages endpoints.
Track usage automatically
Each request is checked against your balance and settled from reported or estimated token usage.
- Estimated maximum request cost is checked before model execution.
- Input and output usage is charged at the published rates.
- Review balance, month-to-date spend, and recent transactions in billing.
Add funds as needed
Use Stripe Checkout to top up the same prepaid balance used by chat and API requests.
- Choose from the available USD top-up amounts.
- Successful payments are credited through the Stripe webhook.
- See completed top-ups and usage charges in recent transactions.
Why adopt Ling 1T
Contrast open-source flexibility with closed platforms
Ling 1T matches or exceeds black-box models on reasoning while remaining auditable, customizable, and cost-effective.
Closed-source offerings
Opaque policies and premium pricing
- Limited visibility into training data and safety guardrails.
- Restrictive usage terms that complicate compliance reviews.
- Inflexible latency and routing tuned for generic workloads.
- Difficult to integrate on-premise due to proprietary tooling.
- High token pricing with unpredictable overage costs.
- Slow iteration cycles when you need custom behaviors.
- Vendor lock-in across SDKs, telemetry, and governance.
- Limited ability to audit reasoning traces or provide citations.
Ling 1T open platform
Transparent, tunable, enterprise-ready
- MIT license with unrestricted self-hosting rights.
- Documented FP8 + MoE stack with reproducible training notes.
- Configurable routing, temperature, and tool policies per domain.
- Deploy with vLLM, SGLang, or custom inference on your GPUs.
- Managed partners offer predictable token-based pricing.
- Evo-CoT and LPO alignment keep answers auditable.
- 128K context and YaRN scaling absorb long records.
- Community contributions accelerate adapters, datasets, and dashboards.
Pay only for what you use
Transparent metered billing: $1.40 per million input compute units and $5.60 per million output compute units.
Ling 1T Pricing
Add funds to your account, then pay for the Ling 1T usage your requests consume. Charges are deducted from your USD balance based on input and output compute units (≈1M tokens).
Input usage
Output usage
Why teams choose Ling 1T metered pricing
- The same rates apply to chat, tool use, and API completions—no plan tiers or commitments.
- Usage charges use reported token counts when available and server-side estimates otherwise.
- The billing dashboard shows your current balance, monthly spend, last top-up, and recent balance transactions.
Cost snapshots
- 0.5M input compute units (~500K tokens) → $0.70.
- 0.25M input + 0.25M output compute units (~250K tokens each) → $1.75.
Prices exclude taxes. Top-ups are charged in USD and usage is deducted from your prepaid balance.
Ling 1T FAQ
Answers to the most common deployment and capability questions
Ready to build with Ling 1T?
Download the weights, connect to a managed provider, or embed Ling 1T into your stack today.