Ling 2.6

Ling Team, Inclusion AI · 2026-06-13 · analysed 2026-08-10 · skill v0.4.0

sparse-MoE efficient-attention text-only open-weights frontier frontier-generalist

Design overview

Ling-2.6-1T is an 80-layer, ~1T-parameter sparse MoE (256 routed experts, 8 active, 1 shared per layer; 4 dense blocks up front) that is remarkable less for its block than for how the block was obtained: it is an architectural transplant of the trained Ling-2.0-1T GQA checkpoint rather than a from-scratch model. A four-step migration (Lightning Attention Conversion, Linear Warmup, MLA Conversion, MLA Warmup) replaced most GQA layers with Lightning Attention — a simple O(N) linear attention — and converted the remaining full-attention layers from GQA to MLA, fusing the old QK Norm into the projections by calibration and adapting the Partial RoPE (rotary dim 64, theta 6M) to TransMLA. The final block mixes 7 linear-attention layers with 1 MLA layer per group of 8 (70:10 over the stack), keeping KV-cache pressure and long-context FLOPs low at the 262,144-token context, which is reached by a 4K-to-256K training curriculum rather than any post-hoc extension. FFNs are fine-grained MoE with grouped routing (8 groups, top-4), sigmoid FP32 scoring, expert bias for auxiliary-loss-free balancing, and a 2.5x output scale. Post-training diverges into two lines from the shared base: Ling-2.6 is an instant model whose Evo-CoT/LPO RL, bidirectional preference alignment, and shortest-correct-response distillation push capability per output token (~4x token efficiency claimed), while Ring-2.6 adds KPop agentic RL (binary-KL masking) and adaptive thinking for long-horizon tool use. Training is FP8 throughout, served by the linghe fused kernels and an AllGather context-parallel scheme, with MTP layers doubling as speculative-decoding drafts.

What the report claims is novel

Atlas assessment: The closest atlas peer is moonshot-kimi-linear-48b-a3b: both are text-only, open-weights hybrids of a linear-attention family with MLA over a fine-grained MoE backbone aimed at agentic/long-context efficiency — and Kimi Linear's own hybrid-ratio ablation explicitly tested and rejected 7:1 in favor of 3:1, which Ling-2.6 adopts on its own scaling-law evidence. The genuinely distinctive axis is that Ling-2.6 is not trained from scratch: it is an architectural transplant of a trained trillion-parameter GQA checkpoint (the QK-Norm absorption via calibration statistics and the Partial-RoPE-decoupled TransMLA conversion are a novel retrofit procedure with no atlas precedent), and it uses Lightning Attention (TransNormerLLM lineage) rather than the delta-rule KDA used by both Moonshot records. The token-efficiency post-training direction (Evo-CoT, LPO, bidirectional preference alignment) and the grouped-routing fine-grained MoE inherited from Ling-2.0 are family-specific rather than novel-at-large; KPop is an incremental refinement of the team's own IcePop (Ring-1T), which is not itself an atlas record. The '1T' scale, ~4x token-efficiency gain, and 87.6 PinchBench lead are the report's own figures and not independently verifiable from the atlas.

Relation to the atlas

Lineage in

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyLing 2.6[1]
OrganisationLing Team, Inclusion AI[2]
Release date2026-06-13[3]
Report URLhttps://arxiv.org/abs/2606.15079[4]
Licensenot disclosed
Open weightsyes[5]

Variants

Scale

Total parameters1T[6]
Active parameters / tokennot disclosed
Layers80[7]
Hidden dimension8192[8]
FFN inner dimension2048[9]
FFN ratio0.25
Query heads64[10]
KV heads64
Head dimension128[11]
Vocabulary size157184[12]
Embedding tyingnot disclosed

Per-variant configs

VariantTotalLayersHiddenFFNQ headsKV headsContext
Ling-2.6-flash104B32409610243232not disclosed
Ling-2.6-1T1T80819220486464262,144
Ring-2.6-1T1T80819220486464262,144

Core block

Block typesparse-MoE
Attention varianthybrid[13]
Attention layer patternlayer groups of M = 8 throughout the 80-layer stack: 7 Lightning Attention (linear) layers + 1 Full Attention (MLA) layer per group, i.e. 70 linear + 10 MLA layers total; the in-group position of the MLA layer is not disclosed[14]
Depth mixingsequential-residual
ActivationSiLU[15]
Expert count256[16]
Experts per token8[17]
Shared experts1[18]
Routinggrouped routing: ngroup = 8 groups with top-4 group selection; router scores with sigmoid in FP32; routed output scaled by 2.5 with normalized top-k probabilities[19]
Load balancingauxiliary-loss-free load balancing with expert bias (bias-update rate gamma = 0.001 during continue pre-training, reduced to 0.0001 for mid-training)[20]
Expert granularityfine-grained MoE inherited from Ling-2.0: 1 shared + 256 routed experts per layer, per-expert intermediate width 2,048 (1,024 for flash)[21]
MethodRoPE[24]
RoPE base frequency6000000[19]
Partial RoPEyes[25]
TypeRMSNorm (epsilon = 1e-6)[26]
Placementnot disclosed
QK-normno[27]
Attention sinksno
Softcappingno

Context

Trained context length262,144[28]
Deployed context length262,144[19]
Extension methodnone — training ran at the full deployed length via a progressive context curriculum (4K -> 32K -> 256K during pre/mid-training; post-training SFT at 256K); no post-hoc extension method used[22]

Tokenizer

Algorithmnot disclosed
Notesnot disclosed

Training

Training tokens9.6T[29]
Data compositionThree-stage mixture. Migration/continue pre-training: ~46% reasoning-intensive domains (math, code), ~50% general corpora (web text), ~4% multilingual; agentic corpus spanning tool use (500+ real-world MCP environments, 3,000+ tools) and agentic coding with bash/web/repository tasks, teacher-generated with verification; long-context corpus (math, web parsing, summarization, RAG fusion, multi-hop reasoning) with rule+model defect detection; web corpus with STEM recall and atomic-fact construction from Wikipedia; multilingual coverage for 21 languages including 1.1T tokens from Fineweb2/Fineweb2-hq and ~70B synthetic web code. Mid-training strips general web data to ~43% general / 42% reasoning / 15% agentic at 256K.[30]
Curriculum / stagingMulti-stage: (1) Migration Pre-Training (~400B tokens, four steps: Lightning Attention Conversion -> Linear Warmup -> MLA Conversion [QK Norm absorption, partial-parameter training, Partial-RoPE-aware TransMLA conversion] -> MLA Warmup); (2) Continue Pre-Training (8T tokens, 4K context, aggressive data-switching strategy adopted over conservative); (3) Mid-Training (~1.2T tokens: 250B at mixed 32K, then 425B at 32K, then 525B at 256K, high-quality mixture). WSM scheduler throughout.[31]
Optimizernot disclosed
LR scheduleWSM scheduler from Ling-2.0: linear warmup to peak learning rate, constant phase until training concludes, final annealing achieved through checkpoint merging[32]
Batch schedulenot disclosed
PrecisionFP8 training from pretraining through supervised fine-tuning, with FP32 optimizer master weights retained for RL initialization; FP8 continuation training for RL; FP32 LM Head in both training and inference[33]
Parallelismco-designed expert, pipeline, and context parallelism with selective activation recomputation; AllGather-based Context Parallel for Lightning Attention (local-recurrence-then-global-correction, no head-divisibility constraint); Triton-fused varlen state-correction kernel (~68% end-to-end speedup at 256K)[34]
Hardwarenot disclosed
Disclosed computenot disclosed

Post-training

SFTyes[35]
Preference optimisationother[36]
Reasoning trainingEvo-CoT RL (from Ling-2.0) with composite rewards: accuracy, formatting penalty for reasoning markers, dynamic length penalty (difficulty-specific), and LLM-judge semantic redundancy penalty; Linguistic Unit Policy Optimization (LPO); Dynamic Pass Rating (DPR) adaptive curriculum; GSPO for token-efficient tool use (process reward on tool-call sequence + zlib-compression repetition penalty); Ring-2.6 adds KPop agentic RL (binary-KL divergence masking) and adaptive thinking with high/xhigh effort modes[37]
Distillationspecialist distillation: reasoning and agentic specialist models are distilled back into the unified Ling-2.6 / Ring-2.6 model (§3.1, §3.2); shortest-correct-response distillation, where proprietary expert models generate SFT responses and only the shortest accurate candidate is retained (§3.1.2)[38]

Modality

Typetext-only[39]
Attachmentn/a[40]

Inference efficiency

KV-cache designhybrid: MLA layers compress the KV cache into a low-rank latent space (KV LoRA Rank 512, Q LoRA Rank 1536, Table 1); Lightning Attention layers use an O(N) linear-attention recurrence with no growing KV cache. The 7:1 mix 'reduces long-context compute cost, KV-cache pressure, and decoding latency'.[41]
Quantisation shippedFP8 training and inference; module-aware FP8 quantization at deployment (attention linears and shared-expert linears in BF16, routed-expert linears blockwise FP8); BF16 inference path also supported; BF16-input/FP32-output router and LM-Head GEMMs[42]
Speculative decodingMTP: two additional MTP layers added during post-training and continued-trained; MTP layers serve as a draft model for speculative decoding (4 speculative steps); parameter sharing across MTP layers with gradient detachment from all but the first (MTP-3-share) raises accepted length to 3.31 vs 2.71 for MTP-1[43]
Serving optimisationslinghe fused-kernel library (open-sourced): QK Norm + RoPE fusion, group RMSNorm + sigmoid-gate fusion, MLA RoPE and Top-K optimization, split-K blockwise FP8 GEMM for small batches; prefix caching and multi-token generation; AllGather context parallelism; ARouter global rollout scheduling with tail-request migration, spillover training-inference overlap, and instance failover; up to 4x prefill/decode acceleration for Ling-2.6-flash vs similar-size peers[44]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
Artificial Analysis Intelligence Index34Abstract, §3.3.1
C-SimpleQA (Correct)76.53Table 5
SimpleQA-Verified (Correct)31.5Table 5
GPQA-Diamond (Mean@4, CoT)76.17Table 5
AIME26 (Mean@64, CoT)87.4Table 5
HMMT-Nov25 (Mean@64, CoT)81.93Table 5
IMO-AnswerBench (Mean@8, CoT)65.81Table 5
LiveCodeBench-v6 (Mean@4)65.58Table 5
bbeh (Pass@1)52.37Table 5
ARCPrize (Mean@4)50.94Table 5
SWE-bench Verified (Claude Code, openhands fc)72.2Table 5
PinchBench (Mean@5)85.24Table 5
ClawEval (Pass@3)51.0Table 5
BFCL-v4 (Accuracy)70.64Table 5
tau2-bench (Mean@4)78.36Table 5
terminal-bench 2.0 (Accuracy)40.45Table 5
IFBench (Mean@5)57.62Table 5
MRCR (16K-256K, Mean)80.37Table 5
LongBenchV2 (Accuracy)48.31Table 5
AIME 2026 (Avg@64) — Ring-2.6-1T xhigh95.78Table 6
LCB-v6 (Avg@4) — Ring-2.6-1T xhigh86.95Table 6
GPQA-Diamond (Avg@16) — Ring-2.6-1T xhigh85.89Table 6
ARC-AGI-2 (Pass@2) — Ring-2.6-1T xhigh66.18Table 6
PinchBench (Avg@3) — Ring-2.6-1T high87.6Table 6
ClawEval (0424, Pass^3) — Ring-2.6-1T high63.82Table 6
SWE-bench Verified (Resolved) — Ring-2.6-1T high74.0Table 6
SWE-bench Pro (Resolved) — Ring-2.6-1T high53.76Table 6
GAIA-2 Search (Pass@1, 3 runs) — Ring-2.6-1T xhigh77.9Table 6
tau2-Average (Acc) — Ring-2.6-1T high84.26Table 6
MMLU (EM) — Ling-2.6-1T-base86.82Table 3
LongBenchv2 (Acc) — Ling-2.6-1T-base43.54Table 3

Provenance references

  1. stated · Title ('Ling and Ring 2.6 Technical Report'); the report covers the Ling-2.6 and Ring-2.6 family
  2. stated · Title byline
  3. stated · Page 2 header: 'arXiv:2606.15079v1 [cs.CL] 13 Jun 2026'; title page also prints 'Date: June, 2026'
  4. stated · pipeline
  5. stated · Abstract: 'We open-source all checkpoints in the 2.6 family'; 'Both base and post-training checkpoints of the 2.6 family are open-sourced'
  6. stated · Abstract ('This model family scales from 104B to 1T parameters'); model name Ling-2.6-1T
  7. stated · §2.1.1 ('the number of Transformer layers to 80'); Table 1 (# Layers 80)
  8. stated · §2.1.1 ('the hidden dimension d to 8,192'); Table 1 (Hidden Size 8,192)
  9. stated · §2.1.1 ('the intermediate hidden dimension of each routed expert is 2,048'); Table 1 (Expert Intermediate Size 2,048)
  10. stated · §2.1.1 ('64 attention heads'); Table 1 (# Attention Heads 64)
  11. stated · §2.1.1 ('64 attention heads with a head dimension of 128'); consistent with 8,192 / 64
  12. stated · §2.1.1 ('a vocabulary size of 157,184')
  13. stated · §2.1 ('a hybrid linear attention design that integrates Lightning Attention with MLA at a 7:1 ratio')
  14. stated · §2.1.2 ('each group contains 1 Full Attention (MLA) layer and M-1 Linear Attention layers'; M = 8 best); Table 1 (Layer Group Size 8)
  15. stated · §2.1.1 ('We employ SiLU as the activation function'); §4.3 calls the fused gated FFN 'SwiGLU' (fc1 -> SiLU gate, multiply with fc2)
  16. stated · §2.1.1 ('256 routed experts'); Table 1 (# Experts (total) 256)
  17. stated · §2.1.1 ('8 experts are activated for each token'); Table 1 (# Experts Active per Token 8)
  18. stated · §2.1.1 ('1 shared expert'); Table 1 (# Shared Experts 1)
  19. stated · §2.1.1
  20. stated · §2.1.1, §2.3.1
  21. stated · §2.1.1 ('Following Ling-2.0, Ling-2.6-1T-base employs a fine-grained MoE architecture'); Table 1
  22. stated · §2.3.2, §3.1.1
  23. n/a · §2.3.2 (no rescaling factor; training at deployed length)
  24. stated · §2.1.1 ('supports a maximum context length of 262,144 tokens via RoPE (Su et al., 2024) positional encoding with theta = 6,000,000')
  25. stated · §2.1.1 ('The rotary dimension is set to 64, i.e., Partial RoPE is applied to a subset of head dimensions.')
  26. stated · §2.1.1 ('RMSNorm with epsilon = 10-6 for layer normalization'); group RMSNorm + sigmoid-gate fusion appears in the inference kernels (§4.3)
  27. stated · §2.1.3
  28. stated · §2.1.1 ('maximum context length of 262,144 tokens'); §2.3.2 (mid-training extends 4K -> 32K -> 256K); §3.1.1 (post-training at 256K)
  29. stated · §2.3 ('processes approximately 9.6T tokens across three stages'); Abstract ('continue pre-training ... for about 9.6T tokens')
  30. stated · §2.2, §2.3.2
  31. stated · §2.3.2
  32. stated · §2.3.1
  33. stated · §4.2
  34. stated · §4.1
  35. stated · §3.1 ('an initial cold-start SFT phase followed by specialized expert fine-tuning'); §3.2 (Ring-2.6 cold-start SFT)
  36. stated · §3.1.3
  37. stated · §3.1.2, §3.1.3, §3.2.2, §3.2.3
  38. stated · §3.1, §3.2
  39. stated · §5 ('an equally important transition is from text-only systems to native multimodal agents')
  40. n/a · §5 (model is text-only)
  41. stated · §2.1, Table 1
  42. stated · §4.2, §4.3
  43. stated · §3.1, Appendix B
  44. stated · §3.3.1, §4.1.1, §4.2, §4.3