Ling 3.0 Flash
Ling 3.0 Flash by InclusionAI: 124B hybrid-linear MoE, 5.1B active per token. KDA+MLA attention, native hybrid reasoning, 10K+ agentic training environments.
Model Overview
Capabilities, design details, and architectural traits
Ling 3.0 Flash - Native Hybrid-Linear Reasoning MoE
Ling 3.0 Flash is a hybrid-linear Mixture-of-Experts model from InclusionAI.
The model's defining architectural choice is a native hybrid-linear attention stack built from the start of pretraining. It alternates Kimi Delta Attention (KDA) layers with Gated Multi-head Latent Attention (MLA) in a 5:1 ratio (35 KDA + 7 Gated MLA + 2 dense layers). KDA introduces fine-grained diagonal gating in Delta Rule state updates, keeping long-context memory stable and predictable. Periodic full-attention MLA layers preserve exact token-to-token recall that pure linear attention loses.
| Trait | Detail |
|---|---|
| Ultra-sparse MoE | 512 routed experts + 1 shared expert; only 8 experts activated per token (1/64 ratio, down from 1/32 in prior generation) |
| Native hybrid reasoning | Combines Flash-series speed with Ring-series deep thinking; thinking mode enabled by default, dynamically scaling effort by task difficulty |
| Agentic training | 10,000+ interactive training environments for end-to-end closed-loop execution across coding, general, and deep research agent tasks |
| Hierarchical caching | Natively integrates SGLang HiCache + Mooncake architecture with physical dual-pools and cluster-shared L3 cache, reducing TTFT by 60-80% in long-input scenarios |
| Context schedule | Trained progressively at 8K -> 32K -> 256K context windows |
Bridging Flash and Ring
InclusionAI's model families serve distinct roles: Ling models prioritize fast, high-throughput production inference, while Ring models specialize in deep step-by-step reasoning. Ling 3.0 Flash fuses both lineages into a single architecture, dynamically switching between rapid non-thinking responses and multi-step reasoning depending on task complexity. This hybrid reasoning mode is the model's core differentiator against pure throughput or pure reasoning models.
Benchmark Performance
Independent evaluations · Artificial Analysis
Accuracy & Capability Details
Compare Models Side-by-Side
Evaluate specifications, pricing, and independent benchmark indices
| Model Details | |||
|---|---|---|---|
| General Info | |||
| Provider | InclusionAI | Anthropic | Anthropic |
| Release Date | August 4, 2026 | September 22, 2026 | September 28, 2026 |
| Knowledge Cutoff | - | - | Jun 2026 |
| Context & Limits | |||
| Context Window | 262K | 1M Best Context Window | 1M Best Context Window |
| Pricing (per 1M tokens) | |||
| Input Pricing | $0.07 Best Input Pricing | $4 | $2 |
| Output Pricing | $0.22 Best Output Pricing | $20 | $10 |
| Modalities | |||
| Inputs | text | textimagefile | textimagefile |
| Outputs | text | text | text |
| Benchmarks (0-100) | |||
| Intelligence Index | 20.1 | 57.6 Best Intelligence Index | 56.0 |
| Coding Index | 50.6 | - | - |
| Agentic Index | 19.3 | - | - |
Humanity's Last Exam
Extremely difficult logical reasoning and knowledge.
Long Context Reasoning
Logical reasoning over long context windows.
SciCode Benchmark
Scientific coding and mathematical modeling.
Independent evaluation data provided by Artificial Analysis. To view the latest benchmarks and full details, visit their official site.
Explore more from InclusionAI
Other models by InclusionAI
Ling 3.1 Flash
Ling-3.0-flash-VL
Ling-3.0-flash-Fin
Ling 3.0 Tiny
Top AI Models
Leading alternatives by intelligence score
