# Ling-2.5-1T (1 Trillion Parameters) Ling-2.5-1T by Ant Group (released February 16, 2026) is one of the largest open-source language models ever released — **1 trillion total parameters with 63B active**. It introduces a hybrid linear attention architecture that enables efficient inference on context lengths up to 1 million tokens. Alongside it, Ant Group released Ring-2.5-1T, the world's first hybrid linear-architecture thinking model. Together, they represent a new frontier in open-source AI — competitive with GPT-5.2, DeepSeek V3.2, and Kimi K2.5 on reasoning and agentic benchmarks. **HuggingFace:** [inclusionAI/Ling-2.5-1T](https://huggingface.co/inclusionAI/Ling-2.5-1T) **Companion model:** [inclusionAI/Ring-2.5-1T](https://huggingface.co/inclusionAI/Ring-2.5-1T) (thinking/reasoning variant) **License:** Open source (Ant Group InclusionAI License) ## Key Features * **1 trillion total parameters, 63B active** — massive scale with efficient MoE-style activation * **Hybrid linear attention** — combines MLA (Multi-head Linear Attention) with Lightning Linear Attention for exceptional throughput on long sequences * **1M token context window** — via YaRN extension from native 256K, handles entire codebases and book-length documents * **Frontier reasoning** — approaches thinking-model performance while using \~4× fewer output tokens * **Agentic capabilities** — trained with Agentic RL, compatible with Claude Code, OpenCode, and OpenClaw * **Ring-2.5-1T companion** — dedicated reasoning variant achieves IMO 2025 and CMO 2025 gold medal level ## Architecture Details | Component | Details | | ----------------- | ------------------------------------------------ | | Total Parameters | 1T (1,000B) | | Active Parameters | 63B | | Architecture | Hybrid Linear Attention (MLA + Lightning Linear) | | Pre-training Data | 29T tokens | | Native Context | 256K tokens | | Extended Context | 1M tokens (YaRN) | | Release Date | February 16, 2026 | ## Requirements Running Ling-2.5-1T at full precision requires substantial resources. Quantized versions make it more accessible. | Configuration | Quantized (Q4 GGUF) | FP8 | BF16 (Full) | | ------------- | ------------------- | -------------- | ---------------- | | GPU | 8× RTX 4090 | 8× H100 80GB | 16× H100 80GB | | VRAM | 8×24GB (192GB) | 8×80GB (640GB) | 16×80GB (1.28TB) | | RAM | 256GB | 512GB | 1TB | | Disk | 600GB | 1.2TB | 2TB+ | | CUDA | 12.0+ | 12.0+ | 12.0+ | **Recommended Clore.ai setup:** * **Quantized (Q4):** 8× RTX 4090 (\~$4–16/day) — usable for experimentation and moderate workloads * **Production (FP8):** 8× H100 (\~$24–48/day) — full quality with good throughput * **Note:** This is an extremely large model. For budget-conscious users, consider the smaller models in the Ling family on [HuggingFace](https://huggingface.co/inclusionAI). ## Quick Start with vLLM vLLM is the recommended serving framework for Ling-2.5-1T: ```bash # Install vLLM pip install vllm # Serve Ling-2.5-1T with tensor parallelism across 8 GPUs vllm serve inclusionAI/Ling-2.5-1T \ --tensor-parallel-size 8 \ --max-model-len 65536 \ --gpu-memory-utilization 0.90 \ --trust-remote-code \ --host 0.0.0.0 \ --port 8000 # For reduced memory, limit context length: vllm serve inclusionAI/Ling-2.5-1T \ --tensor-parallel-size 8 \ --max-model-len 16384 \ --gpu-memory-utilization 0.95 \ --trust-remote-code \ --host 0.0.0.0 \ --port 8000 ``` ## Quick Start with llama.cpp (Quantized) For consumer GPU setups, GGUF quantizations are available: ```bash # Install llama.cpp git clone https://github.com/ggerganov/llama.cpp cd llama.cpp cmake -B build -DGGML_CUDA=ON cmake --build build --config Release -j$(nproc) # Download a quantized GGUF (check HuggingFace for available quants) huggingface-cli download inclusionAI/Ling-2.5-1T-GGUF \ --include "*.Q4_K_M.gguf" \ --local-dir ./models/ # Serve with llama-server (adjust -ngl for your GPU count) ./build/bin/llama-server \ -m ./models/Ling-2.5-1T-Q4_K_M.gguf \ -ngl 99 \ -c 8192 \ --host 0.0.0.0 \ --port 8000 ``` ## Usage Examples ### 1. Chat Completion via OpenAI API Once vLLM or llama-server is running: ```python from openai import OpenAI client = OpenAI( base_url="http://localhost:8000/v1", api_key="not-needed" ) response = client.chat.completions.create( model="inclusionAI/Ling-2.5-1T", messages=[ {"role": "system", "content": "You are a world-class reasoning assistant. Think step by step."}, {"role": "user", "content": "Prove that the square root of 2 is irrational."} ], temperature=0.1, max_tokens=4096 ) print(response.choices[0].message.content) ``` ### 2. Long-Context Document Analysis Ling-2.5-1T's hybrid linear attention makes it exceptionally efficient for long documents: ```python from openai import OpenAI client = OpenAI(base_url="http://localhost:8000/v1", api_key="n/a") # Load a large document with open("full_codebase.txt", "r") as f: codebase = f.read() # Can be hundreds of thousands of tokens response = client.chat.completions.create( model="inclusionAI/Ling-2.5-1T", messages=[ {"role": "system", "content": "You are a senior software architect."}, {"role": "user", "content": f"Analyze this codebase for security vulnerabilities and architectural issues:\n\n{codebase}"} ], temperature=0.1, max_tokens=8192 ) print(response.choices[0].message.content) ``` ### 3. Agentic Tool Use Ling-2.5-1T is trained with Agentic RL for tool calling: ```python from openai import OpenAI client = OpenAI(base_url="http://localhost:8000/v1", api_key="n/a") tools = [ { "type": "function", "function": { "name": "search_database", "description": "Search the product database", "parameters": { "type": "object", "properties": { "query": {"type": "string"}, "category": {"type": "string", "enum": ["electronics", "clothing", "books"]}, "max_price": {"type": "number"} }, "required": ["query"] } } } ] response = client.chat.completions.create( model="inclusionAI/Ling-2.5-1T", messages=[{"role": "user", "content": "Find me a laptop under $1000 with good reviews"}], tools=tools, tool_choice="auto" ) print(response.choices[0].message.tool_calls) ``` ## Ling-2.5-1T vs Ring-2.5-1T | Aspect | Ling-2.5-1T | Ring-2.5-1T | | ---------------- | ----------------------------------- | ---------------------------------------- | | Type | Instant (fast) model | Thinking (reasoning) model | | Architecture | Hybrid Linear Attention | Hybrid Linear Attention | | Best For | General chat, coding, agentic tasks | Math, formal reasoning, complex problems | | Output Style | Direct answers | Chain-of-thought reasoning | | Token Efficiency | High (fewer output tokens) | Uses more tokens for reasoning | | IMO 2025 | Competitive | Gold medal level | ## Tips for Clore.ai Users 1. **This model needs serious hardware** — At 1T parameters, even Q4 quantization requires \~500GB of storage and 192GB+ VRAM. Make sure your Clore.ai instance has sufficient disk and multi-GPU before downloading. 2. **Start with `--max-model-len 8192`** — When first testing, use a short context to verify the model loads and runs correctly. Scale up the context length once everything works. 3. **Use persistent storage** — The model weighs 1–2TB. Attach a large persistent volume on Clore.ai to avoid re-downloading. Download once with `huggingface-cli download`. 4. **Consider Ring-2.5-1T for reasoning tasks** — If your use case is primarily math, logic, or formal reasoning, the companion Ring-2.5-1T model is specifically optimized for chain-of-thought reasoning. 5. **Monitor GPU memory** — With 8-GPU setups, use `nvidia-smi -l 1` to monitor memory usage and watch for OOM during generation with long contexts. ## Troubleshooting | Issue | Solution | | ------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------ | | `CUDA out of memory` | Reduce `--max-model-len`; ensure `--tensor-parallel-size` matches GPU count; try `--gpu-memory-utilization 0.95` | | Very slow generation | Linear attention needs warmup; first few requests may be slow. Also check you have NVLink between GPUs | | Model download fails | Model is \~2TB in BF16. Ensure enough disk space. Use `--resume-download` flag with `huggingface-cli` | | vLLM doesn't support the architecture | Ensure you're using vLLM ≥0.7.0 with `--trust-remote-code`; the custom attention layers require this flag | | GGUF not available | Check [unsloth](https://huggingface.co/unsloth) or community quantizations; the model may take time to be quantized by the community | | Poor quality responses | Use temperature ≤0.1 for factual tasks; add a system prompt; ensure you're not truncating the context | ## Further Reading * [Official Announcement (BusinessWire)](https://www.businesswire.com/news/home/20260215551663/en/) — Release details and benchmarks * [HuggingFace — Ling-2.5-1T](https://huggingface.co/inclusionAI/Ling-2.5-1T) — Model weights and documentation * [HuggingFace — Ring-2.5-1T](https://huggingface.co/inclusionAI/Ring-2.5-1T) — Thinking model companion * [ModelScope Mirror](https://www.modelscope.cn/models/inclusionAI/Ling-2.5-1T) — Faster downloads in Asia * [vLLM Documentation](https://docs.vllm.ai/) — Serving framework