PaperlayerView exact revision on arXiv

Source-linked AI summary

Ling and Ring 2.6 Technical Report: Efficient and Instant Agentic Intelligence at Trillion-Parameter Scale

Ang Li, Ben Liu, Bin Han, Bin Hu, Bin Jing, Binbin Hu, Bing Li, Cai Chen, Caizhi Tang, Changxin Tian, Chao Huang, Chao Zhang, Chen Liang, Chen Qian, Chengfu Tang, Chengyao Wen, Chilin Fu, Chunwei Wu, Cong Zhang, Cunyin Peng, Daixin Wang, Dalong Zhang, Deng Zhao, Dingnan Jin, Dingyuan Zhu, Donghao Zhang, Fan Yuan, Fangzheng Zhao, Fanzhuang Meng, Feifan Wu, Feng Xu, Fengbin Fang, Gangshan Wang, Guodong Yang, Hailin Zhao, Haitao Wang, Haitao Zhang, Hanxiao Zhang, Hanzi Wang, Hao Dai, Hao Liu, Hao Qian, Hao Wu, Haoxiong Liu, Haoyu Xu, Heng Zhang, Hong Liu, Hongliang Zhang, Hongrui Liu, Hongxun Li, Hongzhi Ruan, Huaidong Xiong, Huihuang Zheng, Huikang Tang, Jia Guo, Jia Li, Jia Liu, Jiameng Wang, Jiaming Liu, Jiannan Shi, Jianping Wei, Jiaolong Yang, Jiapeng Wang, Jie Gao, Jie Wang, Jiewei Wu, Jin Yang, Jinjin Li, Jinjing Huang, Jinquan Sun, Jinyao Chen, Juanhui Tu, Jun Liu, Jun Mei, Jun Xu, Jun Zhou, Junjie Ou, Junnan Sipan, Junpeng Fang, Kaihong Zhang, Kaiqin Hu, Ke Shi, Kuan Xu, Kun Tang, Kunlong Chen, Lanyin Mei, Lei Chen, Lei Liang, Lei Xu, Li Tang, Liang Jiang, Liangcheng Fu, Lihui Zhang, Linfeng Shi, Lintao Ma, Liyuan Liu, Longfei Li, Longfei Zheng, Lu Liu, Lu Yu, Man Li, Meiqi Zhu, Meng Li, Mengjie Gao, Mengshu Sun, Mingming Yin, Mingyang Zhang, Mingyuan Fan, Nuo Xu, Pan Tang, Peijie Jiang, Peilong Zhao, Peng Lin, Pingping Liu, Qi Zuo, Qian Zhao, Qiang Cheng, Qianggang Cao, Qiaoben Bao, Qing Cui, Qingyuan Yang, Qitao Shi, Qiyin Huang, Qizheng Zhou, Quan Wan, Runyuan Zhao, Shaomian Zheng, Shaowei Wei, Shengnan Zhang, Shuaicheng Li, Shujie Li, Shuo Zhang, Sikang Bian, Tianchu Yao, Tiange Xu, Tianshu Wang, Ting Guo, Tinghao Wang, Tingwei Huang, Tong Zhao, Tongkai Yang, Wang Hong, Wanli Gu, Wei Lu, Weichang Wu, Weiguang Han, Weiquan Li, Wenbo Shen, Wenjing Fang, Wenzhi Tang, Xiang Shu, Xiao Shi, Xiaodong Yan, Xiaolu Zhang, Xiaopei Wan, Xiaqing Sun, Xin Zhao, Xingyu Lu, Xinxing Yang, Xinyao Tang, Xinyu Kong, Xinyu Liu, Xiong Xu, Xuan Sun, Xudong Han, Xudong Wang, Xujie Shen, Yalin Zhang, Yangyang Hou, Yankun Ren, Yao Zhao, Ye Chen, Yeyang Chen, Yibo Cao, Yifan Zuo, Yijie Chen, Ying Li, Yingjie Song, Yingxue Li, Yiqi Wang, Yixuan Sun, Yizhu Xiao, Yongfei Xu, Yu Liu, Yuchen Fang, Yue Gao, Yue Yu, Yue Zhang, Yuqi Zhang, Yuxiao He, Yuxiao Lu, Yuxin Tian, Yuxuan Li, Yuzhuo Fu, Zhankai Xu, Zhaoxin Huan, Zhenduo Zhang, Zhengke Gui, Zhengyu Huang, Zhenjun Ma, Zhenxuan Pan, Zheping Qu, Zhibo Zhu, Zhidong Fan, Zhigang Huangfu, Zhihao Wang, Zhiqiang Zhang, Zhizhen Liu, Zhuyan Zhou, Zibin Lin, Zihang Zeng, Zihao Wang, Zilong Wang, Ziqi Liu, Zitao Xuan, Zixuan Cheng, Zujie Wen, Zuoli Tang

arXiv:2606.15079v1cs.CLcs.AI

TL;DR

Practical agentic systems must balance reasoning quality, tool reliability, long-context efficiency, latency, and token cost. This report develops Ling-2.6 and Ring-2.6 through architectural migration, token-efficient post-training, and environment-grounded agent training, reporting strong performance across reasoning and agentic benchmarks.

  • Problem

    Practical agentic systems must jointly achieve strong reasoning, reliable tool use, long-context efficiency, and low latency despite tensions among these objectives.

  • Method

    The authors upgrade Ling-2.0 through architectural retrofit, continued pretraining, token-efficient post-training, and environment-grounded agentic reinforcement learning.

  • Results

    Across reasoning and agentic benchmarks, Ling-2.6-1T and Ring-2.6-1T achieve leading or highly competitive performance, including Ring-2.6-1T’s 87.60% on PinchBench.

  • Takeaways & Limitations

    Ling-2.6 and Ring-2.6 provide open model families targeting fast, token-efficient responses and deeper reasoning for practical agentic systems at trillion-parameter scale.

  • Takeaways & Limitations

    Ling-2.6-flash’s tighter deliberation budget constrains reasoning depth, instruction compositionality, and tool-use reliability in high-complexity settings.

Abstract

from arXiv · show

Efficient and scalable agentic intelligence requires models that can deliver both low-latency responses and strong reasoning capabilities while remaining practical to train, serve, and deploy. In this report, we present Ling-2.6 and Ring-2.6, a family of models designed to address this challenge at scale. Ling-2.6 is optimized for instant response generation and high capability per output token, whereas Ring-2.6 is tailored for deeper reasoning and more advanced agentic workflows. Instead of training from scratch, we upgrade the Ling-2.0 base model through architectural migration pre-training and large-scale post-training. This upgrade is guided by a unified co-design of model architecture, optimization objectives, serving systems, and agent training environments, enabling improvements in both model capability and deployment efficiency. At the architectural level, we introduce a hybrid linear attention design that integrates Lightning Attention with MLA, improving the efficiency of long-context training and decoding. To further enhance token efficiency, we optimize capability per output token through Evolutionary Chain-of-Thought, Linguistic Unit Policy Optimization, bidirectional preference alignment, and shortest-correct-response distillation. For agentic capabilities, we propose KPop, a reinforcement learning framework designed to support stable training of Ring-2.6-1T on large-scale environment-grounded data. KPop improves training efficiency through asynchronous scheduling across coding, search, tool use, and workflow execution, enabling scalable learning from complex agent-environment interactions. Together, Ling-2.6 and Ring-2.6 provide a practical pathway toward efficient, scalable, and open agentic systems. We open-source all checkpoints in the 2.6 family to support further research and development in practical agentic intelligence.

1 Introduction

Ling-2.6 and Ring-2.6 target practical agentic intelligence by balancing reasoning quality, response speed, and token efficiency at trillion-parameter scale. The family combines hybrid long-context attention, token-efficient post-training, and native environment-grounded agentic training, while open-sourcing its checkpoints.

  • Introduction: Practical agentic models must reason well, use tools reliably, and remain efficient, but longer reasoning increases latency and token cost while fast models may struggle with reliable long-horizon behavior.The introduction frames these objectives as inherently tensioned in real agent workflows.
  • Introduction: Ling-2.6 prioritizes instant responses and high token efficiency, whereas Ring-2.6 targets deeper reasoning with controllable effort across a 104B-to-1T-parameter family.The report names three open-sourced weights: Ling-2.6-flash, Ling-2.6-1T, and Ring-2.6-1T.
  • Introduction: A unified hybrid linear attention architecture combines Lightning Attention and MLA at a 7:1 ratio to reduce long-context compute cost, KV-cache pressure, and decoding latency.In the earlier GQA-based architecture, attention exceeded 60% of total FLOPs beyond 32K tokens.
  • Introduction: Ling-2.6 optimizes capability per generated token through Evolutionary Chain of Thought, Linguistic Unit Policy Optimization, and bidirectional preference alignment.The stated objective is higher information density while preserving reasoning fidelity, rather than merely shorter responses.
  • Introduction: The family trains agentic abilities directly on tool use, coding, search, workflow execution, and multi-turn interactions with verifiable tasks, structured traces, and environment-grounded feedback.Ring-2.6 additionally introduces KPop, an RL algorithm that replaces IcePop’s uniform fixed-ratio constraint.
  • Introduction: The report presents architecture, post-training, and agentic-training improvements as a path toward practical agentic intelligence and states that base and post-training checkpoints are open-sourced.The claimed improvements address long-context efficiency, intelligence per token, and reliability in real environments.

2 Pre-training

Pre-training upgrades the Ling-2.0-1T checkpoint into a shared Ling-2.6/Ring-2.6 base through a 7:1 Lightning Attention–MLA hybrid, migration strategies, and continued training. The resulting base models improve knowledge, long-context modeling, and reasoning while preserving or further improving mathematics and coding performance.

  • Architecture migration: The models retrofit the Ling-2.0-1T checkpoint rather than training a trillion-parameter replacement from scratch, using a 7:1 Lightning Attention-to-MLA hybrid as the shared family base.This preserves the prior 20Ttoken investment while reducing long-context attention costs from O(n2) to O(n).
  • Hybrid-ratio selection: The M = 8 configuration, corresponding to a 7:1 Linear-to-Full ratio, achieves the best scaling performance and balances model quality with inference efficiency.The authors select this ratio for Ling-2.6-1T-base after equal-FLOPs scaling-law experiments comparing M ∈ {2, 4, 8, 16}.
  • Long-context efficiency: At context windows exceeding 32K tokens, attention becomes the dominant computational bottleneck, and at 256K+ tokens, Full Attention becomes the absolute performance bottleneck.The hybrid architecture addresses this by replacing a subset of GQA layers with Lightning Attention.
  • MLA conversion: Calibration-based fusion removes nonlinear QK Norm while preserving its normalizing effect, enabling subsequent MLA weight absorption for efficient inference.The method approximately fuses QK Norm parameters into query and key projection weights using per-dimension statistics from calibration samples.
  • Evaluation results: Ling-2.6-1T-base models deliver broad gains in knowledge, long-context modeling, and reasoning, while retaining strong mathematics and coding performance without noticeable degradation.Performance further improves on representative benchmarks such as GPQA, SimpleQA, MMMLU, MathBench, and LiveCodeBench.

3 Post-training · 3.1 Post-Training for Ling-2.6

Ling-2.6 and Ring-2.6 share a base model but diverge in optimization targets: Ling-2.6 prioritizes rapid, token-efficient responses and basic agentic capability, while Ring-2.6 emphasizes stronger reasoning and advanced agentic intelligence. Ling-2.6’s post-training uses expert specialization, balanced SFT data, targeted RL, and preference alignment before final distillation.

  • 3 Post-training: Ling-2.6 targets rapid response, high token efficiency, and basic agentic capability, whereas Ring-2.6 emphasizes stronger reasoning and advanced agentic intelligence.
  • 3.1 Post-Training for Ling-2.6: Ling-2.6 replaces Ling-2.0’s unified post-training with expert-driven training: cold-start SFT, specialized expert fine-tuning, specialist RL, and final capability distillation.
  • 3.1.1 Supervised Fine-Tuning Corpus: The SFT foundation balances reasoning, long-context, and agentic tool-use scenarios to support token-efficient responses, deep long-context reasoning, and robust agentic behaviors.
  • 3.1.1 Supervised Fine-Tuning Corpus: The corpus extends the post-training context window to 256K tokens using books, papers, code, financial reports, and web data, with number-dense contexts designed for multi-hop reasoning.
  • 3.1.2 Specialist Training of Ling: Specialist training combines expert SFT and targeted RL to maximize instant-response quality under strict token-efficiency constraints before distillation into Ling-2.6.
  • 3.1.2 Specialist Training of Ling: Ling-2.6’s reasoning RL uses Evo-CoT with accuracy, formatting, dynamic-length, and semantic-redundancy objectives to suppress verbosity while preserving adaptive reasoning.
  • 3.1.2 Specialist Training of Ling: For agentic tasks, GSPO rewards alignment with optimal tool-call sequences and penalizes unnecessary invocations, while Dynamic Pass Rating prioritizes difficult or unstable tasks.
  • 3.1.3 Bidirectional Preference Alignment: Bidirectional preference alignment combines positive incentives and negative penalties for higher-signal preference gradients, with ReportLogic and verification agents evaluating complex writing and instruction-following tasks.

3.2 Post-Training for Ring-2.6 … General-Purpose Agent Tasks

Ring-2.6’s post-training combines specialist data, reproducible execution environments, and agentic reinforcement learning to improve long-horizon coding, search, tool-use, planning, and recovery. Its pipeline uses verifiable task construction and KPop to support scalable agentic RL, raising SWE-bench Verified solve rate from 70.8% to 76.28%.

  • 3.2 Post-Training for Ring-2.6: Ring-2.6 optimizes complex, long-horizon, tool-intensive behavior through specialist training over coding, search, and general tool-use tasks in reproducible environments.The objective includes final-task success, planning, search, tool use, and adaptive interaction under realistic execution constraints.
  • 3.2 Post-Training for Ring-2.6: The tool-use mixture emphasizes repository-level coding, mobile and web information seeking, and workflows requiring planning and recovery from intermediate failures.Across these settings, the data prioritizes verifiability, environmental realism, and interaction diversity.
  • Coding Agent Tasks: Coding agents are trained from approximately 300K GitHub PR-Issue pairs filtered for merged links, test patches, repository size, and benchmark non-overlap.The source spans GH-Archive records from Dec 2015–2023, with repositories required to have >100 stars.
  • Search Agent Tasks: Search data combines stateful synthetic mobile-app universes with constraint-rich web questions, retaining live-tool trajectories that produce correct answers without shortcuts.The mobile environments maintain cross-app consistency and include near-miss entities and numeric fields for aggregation; web instances require retrieval and jointly disambiguating constraints.
  • General-Purpose Agent Tasks: General-purpose agent data covers policy-adherent multi-turn calls, harness-agnostic workflows, 197 validated MCP servers, and over 170 synthetic toolkits.The MCP synthesis spans 12 domains and over 2,400 tools, while the synthetic toolkits provide >2,300 callable functions across search, e-commerce, finance, and scientific computation.
  • General-Purpose Agent Tasks: The coding, search, and general-purpose mixture yields competitive results on SWE-bench, GAIA2 Search, BrowseComp, τ2-bench, PinchBench, and ClawEval.The reported outcomes cover coding, search, and general tool-use capabilities.
  • 3.2 Post-Training for Ring-2.6: KPop replaces IcePop’s uniform ratio constraint with a symmetric binary KL criterion that accounts for token-dependent training-inference mismatch.The symmetric mask requires binary KL divergence to remain small in both directions; with KPop, reward rose from 0.54 to approximately 0.68 and SWE-bench Verified solve rate improved from 70.8% to 76.28%.

3.3 Evaluation

The evaluation examines Ling-2.6 and Ring-2.6 across broad benchmarks spanning knowledge, reasoning, agentic behavior, instruction following, and long-context understanding. It assesses Ling-2.6’s instant-response and token-efficiency goals while comparing Ling-2.6 models with state-of-the-art systems.

  • Evaluation scope: The evaluation covers knowledge, reasoning, agentic behavior, instruction following, and long-context understanding.It is designed to assess overall capability and the distinct optimization targets of the two model families.
  • Model objectives: Ling-2.6 is examined as an instant model optimized for fast response and token efficiency.Ring-2.6 is evaluated as a stronger model, according to the supplied passage.
  • Evaluated models: The Ling-2.6 series evaluation specifically includes high-performance Ling-2.6-1T and lightweight Ling-2.6-flash.Their capabilities are assessed across the five stated benchmark domains.
  • Comparisons: Ling-2.6 models are compared against a wide range of state-of-the-art systems, including leading proprietary models.The supplied passage specifies that these comparisons concern non-reasoning or … systems, but the sentence is incomplete.

Benchmarks

The Ling-2.6 series is evaluated across five domains: knowledge, reasoning, agentic capabilities, instruction following, and long-context understanding. Ling-2.6-1T and Ling-2.6-flash are compared with leading proprietary and open-source models, including instant-response configurations.

  • Evaluation coverage: Evaluations span knowledge, reasoning, agentic capabilities, instruction following, and long-context understanding.The benchmark suite is organized around five domains.
  • Knowledge: Knowledge benchmarks include C-SimpleQA, SimpleQA-Verified, GPQA-Diamond, SuperGPQA, and Humanities-Last-Exam.Reported metrics include Correct, Mean@4 with CoT, EM with CoT, and Mean@4.
  • Reasoning: Reasoning benchmarks include AIME 2026, HMMT, IMO-AnswerBench, LiveCodeBench-v6, bbeh, and ARCPrize.Metrics include Mean@64 with CoT, Mean@8 with CoT, Mean@4, and Pass@1.
  • Agentic capabilities: Agentic evaluations cover SWE-bench Verified, PinchBench, ClawEval, BFCL-V4, τ2-bench, and terminal-bench 2.0.Metrics include Claude Code and openhands fc, Mean@5, Pass@3, Accuracy, and Mean@4 with the gpt-5.2 user model.
  • Instruction following and long-context: Instruction-following and long-context evaluations use IFBench, LIFEBench, LongBenchv2, MRCR, Multichallenge, and Multi-IF.Reported metrics include Mean@5, Accuracy, Mean@16K-256K, and turn-3.
  • Model comparisons: Ling-2.6-1T and Ling-2.6-flash are compared against leading proprietary and open-source models, including Kimi-K2.5, GPT-5.4, and GPT-OSS-120B.The comparison includes non-reasoning or instant-response configurations of the listed proprietary models.

Results

The Ling-2.6 models demonstrate strong performance across knowledge, reasoning, agentic, instruction-following, and long-context benchmarks. Ling-2.6-1T also achieves high token efficiency, while Ling-2.6-flash emphasizes lightweight task execution and inference speed.

  • Knowledge: Ling-2.6-1T achieves strong knowledge performance, scoring 76.53 on C-SimpleQA and 31.50 on SimpleQA-Verified.Ling-2.6-flash leads its comparison group on C-SimpleQA with 60.23.
  • Reasoning: Ling-2.6-1T leads across challenging reasoning benchmarks, including AIME26 (87.40), HMMT-Nov25 (81.93), IMO-AnswerBench (65.81), bbeh (52.37), and ARCPrize (50.94).Ling-2.6-flash provides solid baseline reasoning, including 73.85 on AIME26.
  • Agentic Capabilities: Ling-2.6-1T leads PinchBench (85.24), ClawEval (51.00), BFCL-v4 (70.64), and τ2-bench (78.36) on agentic tasks.Ling-2.6-flash leads its comparison group on SWE-bench Verified (61.20), PinchBench (81.30), and τ2-bench (76.36).
  • Instruction Following and Long-Context: Ling-2.6-1T scores 57.62 on IFBench and 80.37 on MRCR (16K-256K), while Ling-2.6-flash decisively outperforms peers across listed long-context and dialogue benchmarks.Ling-2.6-flash scores 75.93 on MRCR.
  • Token Efficiency: 34 using about 16M output tokens gives Ling-2.6-1T about 4× better token efficiency than Ling-2.0-1T, comparable to GPT-5.4 in the non-reasoning setting.The evaluation attributes this improvement to the post-training recipe.
  • Inference Efficiency: Ling-2.6-flash delivers up to 4× faster serving in both prefill and decode than state-of-the-art models in a similar size regime.The result is attributed to its hybrid attention design and highly sparse MoE architecture.

Benchmarks · Results

Ring-2.6-1T is evaluated across reasoning, OpenClaw, agentic coding, agentic search, and function-calling benchmarks under high- and xhigh-budget inference configurations. It achieves leading or competitive results across these areas, including top open-weights performance on several tasks.

  • Benchmarks: Evaluations span five domains: reasoning, OpenClaw, agentic coding, agentic search, and function calling.The benchmark suite includes AIME 2026, LiveCodeBench-v6, GPQA-Diamond, ARC-AGI-2, HMMT-Feb26, IMO-AnswerBench, PinchBench, ClawEval, SWE-bench, GAIA-2 Search, and τ2-bench.
  • Benchmarks: Ring-2.6-1T is tested in xhigh mode for maximum reasoning depth and high mode for efficiency with reduced reasoning overhead.Comparisons include Kimi-K2.6-Thinking, DeepSeek-V4-Pro, ChatGPT-5.4, Gemini-3.1-Pro, GLM-5.1-Thinking, Claude-Opus-4.7-Thinking, and Claude-Opus-4.6-Thinking under standardized configurations.
  • Results: 95.78% on AIME 2026 places Ring-2.6-1T (xhigh) competitively among top frontier models, while 66.18% on ARC-AGI-2 ranks it first among open-weights models.On LiveCodeBench-v6, it scores 86.95%, ranking second among open-weights models behind Kimi-K2.6-Thinking at 89.80% and ahead of GLM-5.1-Thinking at 82.49%.
  • Results: 87.60% on PinchBench is the highest score among evaluated models, and 63.82% on ClawEval ranks first among open-weights models.The PinchBench result exceeds Gemini-3.1-Pro at 80.00% and ChatGPT-5.4 at 79.95%; ClawEval exceeds Kimi-K2.6-Thinking and GLM-5.1-Thinking, both at 62.30%.
  • Results: 74.00% on SWE-bench Verified and 53.76% on SWE-bench Pro show promising agentic coding capability while remaining below leading comparison models.SWE-bench Verified comparisons include Kimi-K2.6-Thinking at 80.20% and DeepSeek-V4-Pro at 80.60%; SWE-bench Pro compares with GLM-5.1-Thinking at 58.40%.
  • Results: 77.90% on GAIA-2 Search demonstrates strong multi-hop reasoning and information retrieval, competing with DeepSeek-V4-Pro at 78.33%.The score trails ChatGPT-5.4 at 82.71% and Gemini-3.1-Pro at 80.83%, while exceeding GLM-5.1-Thinking at 75.63% and Claude-Opus-4.6-Thinking at 73.75%.
  • Results: 84.26% on τ2-Average is competitive with Gemini-3.1-Pro at 84.42% and Kimi-K2.6-Thinking at 84.28%, while τ2-Telecom reaches 96.71%.These results support robust structured API interactions and multi-turn function calling across the τ2-bench suite.

4 Infrastructure

The 2.6 infrastructure co-design targets throughput, numerical stability, and scheduling efficiency at trillion-parameter scale across long-context training, distributed MoE training, and reinforcement learning. Key mechanisms include scalable Linear Attention context parallelism, phase-specific memory policies, and RL systems for efficient rollout scheduling and bounded policy-version staleness.

  • Long-context training: AllGather context parallelism for Linear Attention combines local recurrence with global correction, removes head-divisibility constraints, and scales linearly with cp_size.Communication is overlapped with local output computation to hide inter-rank costs.
  • Long-context training: 68% end-to-end speedup is achieved at 256K context length through vectorized state correction and consolidated per-sub-sequence output processing.The optimization addresses CPU-side kernel-launch overhead and degraded GPU utilization in variable-length inputs, especially during backward passes.
  • MoE training: Phase-specific memory policies balance throughput and stability for distributed MoE training: aggressive GPU memory use during pre-training and conservative use during long-context post-training.The design co-optimizes expert, pipeline, and context parallelism with selective activation recomputation to manage imbalance and memory volatility.
  • RL infrastructure: ASystem and ARouter improve RL infrastructure by separating control and data flow, scheduling for rollout step completion time, migrating tail requests, and supporting training-inference overlap.ARouter also provides failover with request-level streaming checkpoints, avoiding full-step reruns for long-running rollout steps.
  • RL infrastructure: A dedicated staleness manager bounds policy-version skew by admitting segments only within max_staleness × consumer_batch_size and discarding or retiring older segments.Each rollout segment is tagged with the inference policy version used during generation.

5 Conclusion, Limitations, and Future Directions

The report presents Ling-2.6 and Ring-2.6 as a trillion-parameter model family whose progress depends on co-design across architecture, post-training, systems, and agent training environments. It also identifies unresolved reasoning and reliability bottlenecks and calls for continued efficiency co-design and broader capability ceilings.

  • Conclusion: Ling-2.6 and Ring-2.6 target practical agentic intelligence at trillion-parameter scale through co-design across architecture, post-training, systems, and agent training environments.The report emphasizes efficient long-context processing, higher capability per output token, and more reliable environment-grounded capabilities.
  • Limitations: Ling-2.6-flash’s throughput and token economy come with reduced reasoning depth, instruction compositionality, and tool-use reliability in high-complexity settings.The limitation is attributed to its tighter deliberation budget.
  • Limitations: The current token-efficiency objective compresses procedural reasoning effectively but does not always separate low-value repetition from necessary factual elaboration.The passage identifies the token-efficiency objective as incomplete.
  • Future Directions: Future progress requires deeper efficiency co-design across architecture and systems, including model scaling, advanced designs, low-precision training and inference, KV-cache management, and Muon optimization recipes.The report frames this as part of pushing the Ling and Ring family toward higher capability ceilings.

6 Contributors

This section lists the paper’s contributors alphabetically by first name. A dagger marks corresponding authors.

  • Contributors are listed alphabetically by first name.
  • The alphabetical list begins with Ang Li and continues through Hao Wu.
  • The contributor list continues from Haoxiong Liu through Shaowei Wei.
  • The final contributors range from Shuaicheng Li through Zuoli Tang, with daggers identifying corresponding authors.The dagger notation appears next to Jun Zhou and Zhiqiang Zhang, and is defined as denoting corresponding authors.

A Agentic Coding Environments

Agentic coding tasks require isolated, reproducible execution environments for file and system-utility interactions, training, and evaluation. The paper adopts a Claude Code + MCP workflow to automatically generate Docker images by exploring repositories and documentation to produce working Dockerfiles.

  • Environment requirements: Agentic coding tasks run in isolated environments where agents interact with files and system utilities to solve problems.
  • Environment requirements: Reproducible execution environments support faithful rollout and verification during both training and evaluation.
  • Environment construction: The adopted Claude Code + MCP approach automatically generates Docker images for each coding-task instance.Claude Code explores the repository’s codebase and documentation, gathers context, and produces a working Dockerfile.

B Multi-Token Prediction with Continued Training

Ling-2.6 improves inference efficiency during continued training by using multi-token prediction (MTP), which can enhance base-model performance and support speculative decoding. To reduce training–inference discrepancy, the method adds and continues training two MTP layers, producing moderate accepted-length gains while revealing that later layers may remain insufficiently trained.

  • Method: Multi-token prediction improves base-model performance and can serve as a speculative-decoding draft model to accelerate inference.MTP is introduced as an inference-efficiency improvement during continued training.
  • Method: Standard MTP predicts only the next token during training, creating a training–inference discrepancy when multiple tokens are predicted at inference.The discrepancy reduces accepted length, motivating additional MTP layers during training.
  • Method: The post-training procedure adds two MTP layers and continues training the MTP layers to alleviate the training–inference discrepancy.This design directly targets the mismatch caused by using multiple-token prediction at inference.
  • Results: Under four speculative steps, continued training yields a moderate accepted-length improvement over the standard MTP model, while using only the first layer produces a notable increase.The results suggest that the newly introduced MTP layers may remain insufficiently trained in the continued-training setting.
Loading 2606.15079v1…