Architecture diagram of Ling 2.5 (1T) — Figure 32: Ling 2.5 1T · diagram by Sebastian Raschka · click to zoom

Sparse hybridInclusionAI 🇨🇳 · February 15, 2026

Ling 2.5 (1T)

Trillion-parameter long-context model that swaps DeltaNet for Lightning Attention.

Key detail

Uses a 7:1 linear-attention/MLA ratio and a much larger 63B active path.

Total params

1T

Active / token

63B

Context

256K

Fact sheet

Scale: 1T total, 63B active (6.3% active)
Context window: 256,000 tokens
Decoder type: Sparse hybrid
Attention: Lightning Attention plus MLA
Layer mix: 10 MLA + 70 Lightning Attention
KV cache / token (bf16): 11.2 KiB · Very low
License: MIT License
Released: February 15, 2026

Parameters active per token

6.3%

63B of 1T weights run for each token.

KV cache per token (bf16)

11.2 KiB

Smaller than 88% of models in the atlas. How it’s calculated

Architecture concepts

Sources & resources

View in article config.json Original gallery card

Compare with… JSON

Use the ← and → keys to step through releases.

Similar architectures

Ling 2.6 (1T)
InclusionAI, released 2026-04-23. Hybrid MoE decoder, 1T total, 63B active, 256K context. Attention: Lightning Attention plus MLA.
Ling 3.1 Flash (560B)
InclusionAI, released 2026-10-13. Hybrid MoE decoder, 560B total, 25B active, 1M context. Attention: 7:1 Kimi Delta Attention and gated MLA.
Kimi K3 (2.8T)
Moonshot AI, released 2026-07-27. Hybrid MoE decoder, 2.8T total, 104B active, 1M context. Attention: 3:1 Kimi Delta Attention and Gated MLA with NoPE.
Ember-1 (2.8T)
Fireworks AI, released 2026-09-23. Hybrid MoE decoder, 2.8T total, 104B active, 1M context. Attention: 3:1 Kimi Delta Attention and Gated MLA with NoPE.