Recently, we successively released and open-sourced the Bailin 2.6 series models, including Ling-2.6-flash, Ling-2.6-1T, and Ring-2.6-1T.
Today, we officially release the Ling & Ring 2.6 Technical Report, systematically disclosing the technical details of the Bailin 2.6 series across model architecture, pre-training, post-training, Agent reinforcement learning, and inference infrastructure.
The Bailin 2.6 series targets an evolving LLM usage scenario: models are no longer just chat systems answering questions but are progressively entering real-world complex tasks such as Agents, Coding, scientific research analysis, and enterprise workflows. In these scenarios, models need to simultaneously possess three capabilities: reliable reasoning, stable tool use, and the ability to continuously execute tasks under controllable cost and latency.
Towards this goal, we designed the Bailin 2.6 series as a model family oriented towards different task complexities:
Ling-2.6 targets instant response and high token efficiency, focusing on improving capability density per output token;
Ring-2.6 targets deeper reasoning and complex Agent workflows, focusing on improving long-horizon planning, tool calling, code execution, search, and environment interaction capabilities.
Meanwhile, the base and post-training checkpoints of the Bailin 2.6 series have been open-sourced to the community. We aim to provide developers and researchers with a reproducible, extensible, and continuously evolvable technical foundation for Agentic Intelligence by opening up the models, technical report, and related toolchains.
Key Features
The Bailin 2.6 series is not merely a parameter scale upgrade. We focus more on: at the trillion-parameter scale, how to make models more efficient, stable, and usable in real workflows.
Around this goal, Bailin 2.6 underwent systematic optimization in three main directions.
More Efficient Long-Context Capability
Long context only possesses real application value when it is sufficiently efficient. Previously, GQA-based architectures saw attention computation gradually become the primary bottleneck once context length exceeded 32K tokens. To address this, Ling / Ring 2.6 adopts a unified Hybrid Linear Attention architecture, combining Lightning Attention with MLA at a 7:1 ratio, reducing long-context training, decoding, and KV Cache costs while preserving modeling quality as much as possible.
At the system level, we further combined continued MTP training, optimized context parallel communication, and the linghe fused kernel library, enabling the new architecture to achieve better throughput in long-output and long-context training scenarios. Notably, Ling-2.6-flash achieves 340 tokens/s on 4×H20 hardware.
Higher Per-Token Capability Density
For Ling-2.6, token efficiency is a core optimization target. We do not view shorter output as a superficial style optimization; rather, we want the model to carry higher information density in fewer output tokens while maintaining answer quality by reducing redundant reasoning.
In the post-training phase, we combined Evolutionary Chain of Thought (Evo-CoT), Linguistic Unit Policy Optimization (LPO), bidirectional preference alignment, and shortest-correct-response distillation to enhance the model's ability to select effective reasoning steps and reduce repetitive, circular, and low-information-density output.
On the Artificial Analysis Intelligence Index, Ling-2.6-1T scored 34 using approximately 16M output tokens, achieving roughly a 4x improvement in token efficiency on reasoning workloads compared to Ling-2.0-1T.
Natively Optimized Agent Capability
The Agent capability of the Bailin 2.6 series is not simply "transferred" from general chat data but is optimized as a direct training objective.
We constructed a large-scale Agentic Corpus covering tool calling, code, search, workflow execution, and multi-turn interaction, combining these data with verifiable tasks, structured tool trajectories, and environmental feedback. On Ring-2.6, we further propose KPop, replacing the uniform fixed-ratio constraint in IcePop with binary KL divergence to more stably conduct Agentic RL training for MoE models. Simultaneously, we adopt asynchronous RL, decoupling rollout collection from parameter updates, enabling more efficient training of long-horizon tasks like coding, search, tool calling, and workflow execution at the trillion-parameter scale.
These designs enable Ring-2.6-1T to demonstrate stable performance across multiple real-world tasks and Agent benchmarks. For example, Ring-2.6-1T high scores 87.60 on PinchBench, 63.82 on ClawEval, and shows strong multi-step task execution and tool-calling capabilities on GAIA-2 Search and τ2-Bench Telecom.
Note: The upper chart shows Ling-2.6-1T's token efficiency on the Artificial Analysis Intelligence Index; the lower chart shows Ring-2.6-1T's performance on complex reasoning and Agent task benchmarks.
Model Positioning: Ling for Efficiency, Ring for Deep Reasoning and Complex Execution
As LLMs evolve from Chatbots to Agentic Systems, optimization objectives shift. A practical large model must not only possess reasoning ability but also reliably call tools and maintain stable execution in real tasks. Simultaneously, the model must balance response speed, token cost, and task completion quality.
Therefore, we did not attempt to cover all tasks with a single model but designed the Bailin 2.6 series as a model family for different usage scenarios.
Ling-2.6-flash targets low latency, high throughput, and high-frequency invocation, suitable for information extraction, format conversion, long output, batch processing, and lightweight execution nodes in Agent workflows.
Ling-2.6-1T targets higher capability density and stronger general capabilities, emphasizing completing high-quality tasks with fewer output tokens in instant / non-reasoning scenarios.
Ring-2.6-1T targets complex reasoning and long-horizon Agent workflows, supporting high and xhigh reasoning configurations. High is better suited for high-frequency Agent workflows and production default calls; xhigh is for complex reasoning tasks requiring larger thinking budgets.
The core of the 2.6 series division of labor is to let developers choose the more appropriate model and inference configuration based on task complexity, cost constraints, and latency requirements.
Pre-training: Starting from Ling-2.0 Base, Completing Long-Context Architecture Migration
Retraining a model from scratch at the trillion-parameter scale is extremely costly and would lose the massive training gains accumulated by the previous generation. Therefore, Ling / Ring 2.6 did not train from zero but performed architecture migration, continued pre-training, and large-scale post-training on the Ling-2.0 base.
Hybrid Linear Attention: Balancing Long-Context Efficiency and Model Quality
In long-context tasks, traditional Full Attention's computation and cache overhead rise rapidly with context length. To solve this, we adopt a hybrid linear attention architecture combining Lightning Attention and MLA.
Lightning Attention reduces computational complexity along the sequence dimension from O(n²) to O(n), making it more suitable for long-context training and decoding; MLA compresses KV Cache via low-rank latent space, reducing memory pressure during long-context inference.
We compared multiple mixing ratios via scaling law experiments, including 1:1, 3:1, 7:1, and 15:1. Results show the 7:1 Linear-to-Full ratio achieves a better balance between model quality and inference cost, becoming the final architecture choice for Ling-2.6-1T-base.
Note: Scaling law comparison under different Hybrid Attention ratios; 7:1 achieves a better balance between quality and efficiency.
Four-Stage Architecture Migration: Avoiding Training from Scratch, Smoothly Inheriting Existing Capabilities
Migrating from GQA-based Softmax Attention to Hybrid Linear Attention + MLA is not a simple structural replacement. We adopted a four-stage migration process:
- Stage 1: Convert part of GQA layers to Lightning Attention;
- Stage 2: Align new parameters via Linear Warmup;
- Stage 3: Complete MLA Conversion, including QK Norm removal and Partial RoPE adaptation;
- Stage 4: Restore loss to pre-migration levels via MLA Warmup.
This migration stage used approximately 400B tokens, allowing the model to retain Ling-2.0's existing capabilities while gradually adapting to the new efficient long-context architecture.
9.6T Tokens Continuous Training: From 4K to 256K Context
After completing architecture migration, we continued large-scale full-parameter training. Ling-2.6 pre-training processed approximately 9.6T tokens total, divided into three phases:
- Migration Pre-Training: ~400B tokens for architecture migration;
- Continue Pre-Training: ~8T tokens, 4K context, full-parameter continued training;
- Mid-Training: ~1.2T tokens, gradually extending context window from 4K to 32K, then to 256K.
In data composition, we focused on enhancing mathematics, code, Agentic Data, long-context corpora, and multilingual corpora. The Agentic Corpus covers 500+ real MCP environments, 3000+ tools, and various coding, bash, web QA, and software repository tasks; the Long-Context Corpus covers mathematics, complex web parsing, long-document summarization, RAG fusion, and multi-hop reasoning tasks.
Note: Ling-2.6 multi-stage pre-training flow: from architecture migration, continued pre-training to long-context mid-training, gradually extending to 256K context.
Base Model Evaluation: Simultaneous Improvement in Knowledge, Long-Context, Reasoning, and Code Capabilities
In base model evaluation, we used 31 benchmarks covering mathematics, code, general reasoning, language understanding, world knowledge, and long-context understanding to compare Ling-2.6-flash-base, Ling-2.6-1T-base with the 2.0 generation models.
Overall, Ling-2.6-1T-base achieves stable improvements in world knowledge, long-context modeling, and reasoning capabilities while maintaining mathematics and code abilities. Improvements are particularly notable on knowledge and long-context tasks like SimpleQA, C-SimpleQA, MMMLU, and LongBenchv2. This indicates the new high-quality data and long-context training process effectively enhanced the model's knowledge expression, long-range dependency modeling, and multi-step reasoning foundational capabilities.
Note: Comparison of Ling-2.6-base vs Ling-2.0-base on knowledge, mathematics, code, reasoning, language understanding, and long-context benchmarks.
Ling-2.6 Post-Training: Expert-Driven Training Centered on Token Efficiency
Ling-2.6's post-training goal is to obtain a model better suited for instant response and high-frequency invocation. Therefore, we care not only whether the model answers correctly but also whether it can deliver higher-quality answers with fewer output tokens.
Unlike Ling-2.0's unified post-training, Ling-2.6 adopts an expert-driven training paradigm: first cold-start SFT, then expert training for reasoning specialist and agentic specialist; subsequently reinforcing each expert model via RL, finally distilling expert capabilities back into the unified Ling-2.6 model.
Note: Ling-2.6 post-training flow: Cold-start SFT, expert training, token-efficient RL, bidirectional preference alignment, and capability distillation.
Less Redundancy, Higher Information Density
In the reasoning direction, we first used expert models to generate candidate responses and retained the shortest correct answer. Simultaneously, we used an LLM judge to remove over-reasoning segments where the model "continues reflecting after already finding the correct answer." This data-level processing reduced average output length by approximately 200 to 300 tokens.
In the RL phase, we introduced redundancy penalties based on Evo-CoT, including dynamic length penalty and semantic redundancy penalty. Dynamic length penalty allows the model to reason more thoroughly on difficult tasks but limits overlong output on simple tasks; semantic redundancy penalty suppresses circular, repetitive, and low-value reflection by segment-wise evaluation of the reasoning process.
Token-Efficient Agent Training for Tool Use
In Agent tasks, we adopt GSPO and introduce two reward signals: consistency between tool-call trajectories and optimal tool-call sequences, and a repetition penalty based on zlib compression ratio.
The intuition behind compression ratio penalty is simple: highly repetitive and degenerate outputs compress more easily, thus receiving stronger penalties. This makes the model favor concise, coherent, and effective execution paths during tool use.
Simultaneously, we introduce Dynamic Pass Rating (DPR) for dynamic sample selection. Tasks stably solved early in training are treated as easier; tasks unsolved long-term or performing unstably are prioritized for subsequent training. This concentrates training resources on high-information samples near the model's current capability boundary.
Bidirectional Preference Alignment: Balancing Quality and Conciseness
Ling-2.6's final stage introduces bidirectional preference alignment. Unlike single-direction reward modeling, we unify positive incentives and negative penalties into one reward model: encouraging informative, constraint-satisfying answers while penalizing logical errors, hallucinations, and mechanically verbose output.
To prevent the model from gaming reward by simply increasing output length, we further use focus reward, dynamically adjusting training weights based on saturation levels across dimensions, shifting optimization focus toward dimensions still needing improvement.
Ring-2.6 Post-Training: Reinforcement Learning for Long-Horizon Agents
Ring-2.6's post-training objective differs from Ling-2.6. It focuses more on complex, long-horizon, tool-intensive Agent behavior. We want Ring-2.6 not only to improve final task success rates but also to enhance planning, search, tool calling, and adaptive interaction capabilities under real execution constraints.
In training flow, Ring-2.6 starts from Ling-2.6-1T Base, undergoes cold-start SFT, then enters reasoning & agentic experts trained by the KPop algorithm, specialist distillation, finally forming high and xhigh reasoning configurations.
High balances reasoning depth and response conciseness via moderate length penalty; xhigh uses smaller length penalty, releasing more thinking budget for complex reasoning tasks.
Note: Ring-2.6 post-training flow: from Ling-2.6-1T Base to high / xhigh Adaptive Thinking.
Agentic Data: From Code and Search to General Tool Workflows
Ring-2.6's tool-use data mainly covers three capabilities: repository-level coding, mobile/web search, and general tool workflows requiring multi-step planning and error recovery.
In Coding Agent tasks, we mined PR-Issue pairs from GitHub at scale, retaining tasks with stars >100, merged PRs linked to closed issues, and containing test patches, yielding ~300K raw pairs. To prevent overlap with existing SWE benchmarks, we exclude relevant repositories, reducing data contamination risk.
In Search Agent tasks, we built mobile app search and web search environments. Mobile tasks simulate stateful apps like contacts, messages, email, calendar, shopping, travel, files; web search tasks construct multi-hop QA from Wikipedia evidence paths, expressing key facts indirectly to avoid keyword shortcuts.
For general tool use, we built data covering business policy constraints, multi-turn tool calling, harness-agnostic workflows, large-scale MCP synthetic tasks, and general tool calling tasks. The large-scale MCP synthetic tasks cover 197 verified MCP servers, 12 domains, and 2400+ tools.
Agentic RL: Training Models in Real Environmental Feedback
In the Agentic RL phase, we built a lightweight Agent framework providing execute_bash, search_replace, and task_done core tools. Maximum dialogue length is 200 turns during training, 500 turns during evaluation.
For SWE-type long-horizon tasks, we conduct RL training in sandbox environments. These tasks often require 30 to 200 solution steps, needing reproducible environments, reliable verification signals, and high training stability.
The final training dataset contains ~2500 instances from 1550 repositories, covering 30+ programming languages including Python, Java, C, Rust, JavaScript. To reduce reward hacking, we limited git history access and identified potential cheating trajectories via real-time monitoring. Analysis shows ~0.2% of trajectories exhibit cheating patterns, with minimal practical impact.
KPop: Mitigating Train-Inference Inconsistency in Reinforcement Learning
In Ring-1T, we proposed IcePop, improving MoE model RL stability via double-sided masking. However, in Ring-2.6 training, we observed that fixed-ratio constraints implicitly assume all tokens have identical noise structure, which doesn't match real training processes.
Different tokens have different probabilities; the mismatch between training and inference policies also differs. Fixed-ratio masking tends to over-mask low-probability tokens.
Therefore, we propose KPop, replacing IcePop's fixed-ratio constraint with a symmetric binary KL criterion. For each output token, KPop treats the entire vocabulary as two events: "currently sampled token" and "all other tokens," computing binary KL divergence between training and inference policies in both directions. Only when both directional binary KLs are sufficiently small is the token retained for policy update.
Experiments show KPop enables the Agentic RL reward curve on coding tasks to rise continuously during training, improving the lightweight agent's solve rate on SWE-bench Verified from 70.8% to 76.28% (average of three independent runs).
Note: KPop's Agentic RL training dynamics on coding tasks; reward curve rises continuously with training.
Note: KPop's evaluation results on SWE-bench Verified.
Evaluation Results: Ling Faster and More Practical, Ring Better at Complex Reasoning and Agent Execution
We systematically evaluated Ling-2.6 and Ring-2.6 respectively.
- Ling-2.6 is evaluated primarily as an instant model, focusing on fast response, token efficiency, instruction following, and long-context capability;
- Ring-2.6 is evaluated as a long-horizon agentic model, focusing on complex reasoning, OpenClaw, Agentic Coding, Agentic Search, and Function Calling.
Ling-2.6: General Capabilities Under High Token Efficiency
Ling-2.6-1T performs stably across knowledge, reasoning, Agent, instruction following, and long-context tasks.
- On knowledge tasks, Ling-2.6-1T scores 76.53 on C-SimpleQA and 31.50 on SimpleQA-Verified; on reasoning tasks, AIME26 87.40, HMMT-Nov25 81.93, IMO-AnswerBench 65.81, ARCPrize 50.94.
- On Agent tasks, Ling-2.6-1T shows strong tool use and task execution on PinchBench, ClawEval, BFCL-v4, and τ2-bench. Ling-2.6-flash, though a lightweight model, also demonstrates good execution on SWE-bench Verified, PinchBench, TAU2-Bench.
- On long-context and multi-turn dialogue, Ling-2.6-1T scores 80.37 on MRCR 16K-256K; Ling-2.6-flash scores 75.93 on MRCR and performs stably on Multichallenge and Multi-IF. This indicates Ling-2.6-flash possesses good deployment efficiency and task performance in long-input, long-output, and complex interaction scenarios.
Note: Evaluation results of Ling-2.6-flash and Ling-2.6-1T on knowledge, reasoning, Agent, instruction following, and long-context benchmarks.
Ling-2.6-flash: Lightweight Model for Deployment Efficiency
Ling-2.6-1T targets higher capability ceiling, while Ling-2.6-flash emphasizes inference efficiency at deployment. Benefiting from the hybrid attention architecture and highly sparse MoE design, Ling-2.6-flash achieves higher serving speed among similar-scale models.
Under 4×H20 deployment, batch size 32, 4 tensor parallel ranks, output length 64K, Ling-2.6-flash's decode throughput reaches 1.3x Nemotron-3-Super, 2.4x Qwen3.5-122B-A10B, 4.3x GLM-4.5-Air. For long-output, high-frequency, and throughput-constrained Agent workloads, response speed itself is a critical component of model usability.
Note: Ling-2.6-flash throughput performance in prefill and decode stages.
Ring-2.6: Complex Reasoning and Agent Execution Capability
Ring-2.6-1T provides high and xhigh reasoning configurations. Xhigh uses larger thinking budget to maximize complex reasoning depth; high reduces reasoning overhead, better suited for Agent execution and high-frequency invocation.
- On complex reasoning tasks, Ring-2.6-1T xhigh scores 95.78 on AIME 2026, 86.95 on LiveCodeBench-v6, 66.18 on ARC-AGI-2. In the report, Ring-2.6-1T xhigh is the top open-weights model on ARC-AGI-2, demonstrating strong abstract reasoning and pattern recognition capabilities.
- On OpenClaw-related evaluations, Ring-2.6-1T high stands out: PinchBench 87.60, ranking first among listed models; ClawEval 63.82, ranking first among open-weights models, surpassing Kimi-K2.6-Thinking and GLM-5.1-Thinking.
- On Agentic Coding, Ring-2.6-1T high scores 74.00 on SWE-bench Verified and 53.76 on SWE-bench Pro, indicating the model has foundational capabilities for multi-file reasoning, real software engineering tasks, and iterative debugging.
- On Agentic Search, Ring-2.6-1T xhigh scores 77.90 on GAIA-2 Search, showing multi-hop search and information integration capability. On Function Calling, Ring-2.6-1T high scores 84.26 on τ2-bench Average and 96.71 on τ2-Telecom, indicating good reliability in structured API interaction and multi-turn function calling tasks.
Note: Ring-2.6-1T evaluation performance on complex reasoning, OpenClaw, Agentic Coding, Agentic Search, and Function Calling.
Infrastructure: System Coordination Across Training, RL, and Inference
For trillion-parameter models, the system stack itself is a critical component of model capability and cost. Bailin 2.6's infrastructure optimization revolves around three layers: long-context training efficiency, large-scale asynchronous Agentic RL efficiency, and inference serving efficiency.
Long-Context Training: Context Parallel for Lightning Attention
After introducing Hybrid Linear Attention, long-context training is no longer a simple extension of the Ling-2.0 training stack. Lightning Attention has recursive dependencies along the sequence dimension, preventing direct application of traditional Softmax Attention's Context Parallel schemes.
We propose AllGather-based CP, using a "local recursion + global correction" approach. Each rank first computes hidden states and outputs on its local sequence shard, then aggregates states via AllGather for correction. This design doesn't rely on head-count divisibility constraints, better suiting ultra-long-context training.
Simultaneously, we use Triton fused kernels for varlen sequences, unfolding state correction recursion into vectorized matrix form and merging output correction for multiple sub-sequences. At 256K context length, this optimization yields ~68% end-to-end speedup.
Note: Lightning Attention Context Parallel optimization.
ASystem and ARouter: Supporting Large-Scale Asynchronous Agentic RL
Ling / Ring 2.6's RL training is built on ASystem. ASystem is a hybrid SingleController-SPD runtime separating control flow from data flow, treating training, inference, and reward computation as pluggable backends.
In RL rollout, a few long-tail decoding requests can stall the entire batch, leaving GPUs waiting. ARouter's goal isn't minimizing single-request latency but minimizing the entire RL step completion time.
ARouter tracks inference instance load and generation progress, migrating late long-tail requests from congested instances to idle ones; it also supports spillover-based training-inference overlap, letting the main inference group release compute resources and enter training-side gradient accumulation. In long-sequence scenarios, this mechanism delivers over 80% end-to-end performance improvement.
Note: ARouter architecture: global scheduling and long-tail request handling for RL rollout.
linghe: Consistency Optimization from Training to Inference
On the inference side, we further adapt fused kernels accumulated during training to real deployment scenarios, preserving numerical behavior consistency between training and inference as much as possible. This not only improves inference efficiency but also helps reduce train-inference discrepancies in RL rollout.
In BF16 inference, we fuse QK Norm + RoPE, Group RMSNorm + Sigmoid Gate, and optimize MLA RoPE and Top-K; in FP8 inference, we further fuse RMSNorm, SwiGLU, and quantization, introducing Split-K Blockwise FP8 GEMM to improve small-batch throughput.
Combined with kernel fusion, prefix caching, and multi-token generation, linghe brings not only overall throughput gains but also higher per-user TPS, shorter wait times, and more stable interactive experience.
Note: linghe inference operator fusion optimization in Attention and MoE modules.
Note: Throughput improvement from MTP and linghe under different precision and batch size settings.
Limitations and Future Directions
Our current models still have several shortcomings.
- First, while Ling-2.6-flash achieves significant gains in throughput and token economy, its tighter thinking budget limits reasoning depth, complex instruction composition capability, and tool-calling reliability on high-complexity tasks.
- Second, the token efficiency objective still has room for improvement. It effectively compresses redundancy in programmatic reasoning but sometimes cannot fully distinguish low-value repetition from necessary fact elaboration in knowledge-intensive output.
- Third, long-horizon Agent robustness remains weaker than short-horizon task capability. The model may perform well in local decisions but reliability degrades in longer workflows, continuously changing tool states, and heterogeneous execution environments.
- Finally, existing public benchmarks still cannot adequately measure persistence, recovery behavior, cost-aware planning, and deployment-time robustness—key capabilities in real production environments.
Next, we will continue advancing along the co-design direction: deeper coordination among model architecture, systems, low-precision training and inference, KV Cache management, optimization methods, and inference efficiency; simultaneously, we will push Ling / Ring from text-only systems toward native multimodal agents, enabling models to more naturally handle visual interfaces, documents, code, and mixed-modality environments.
Making Agentic Intelligence More Efficient, Open, and Deployable
The core message of the Ling / Ring 2.6 Technical Report is: trillion-parameter model progress should not come from model scale alone, but from system coordination across architecture, post-training, infrastructure, and Agent training environments.
Ling-2.6 focuses more on instant response, token efficiency, and broad usability under deployment constraints; Ring-2.6 focuses more on complex reasoning, long-horizon tasks, and Agent execution in real environments. Together, they form the Bailin 2.6 series' technical path toward practical Agentic Intelligence: long context must be efficient, output tokens must carry higher information density, and Agent behavior must be stably optimized in real environmental feedback.
With the open-sourcing of the 2.6 series models, base, and post-training checkpoints, we hope to open this path to more researchers and developers, jointly exploring efficient, open, and deployable Agent systems.