Model

Ling-3.0-flash

OPEN

inclusionAI · Ling 3.0 family · released Jul 24, 2026

inclusionAI's Ling 3.0 mid-tier, launched 2026-07-24 and open-sourced on Hugging Face under MIT on 2026-08-02. It is a 124B-total / 5.1B-active hybrid-linear MoE with a native 256K context (extendable to 1M) and thinking on by default but switchable. Ant says it matches or beats its predecessor Ring-2.6-1T on key benchmarks, but its comparison charts are images only, so no table values can be transcribed. The one text-stated figure is MiniAppBench 25.3% (in-house), which has no catalog slug.

ReasoningCodingVisionFunction callingTool useAgentic

—

Elo · unrated

Parameters

124B

Active params

5.1B (MoE)

Context

256K tokens

Architecture

Hybrid-linear MoE: 35 KDA + 7 Gated MLA layers (5:1), 2 dense layers, 512 routed / 1 shared / 8 active experts, hidden 2560

License

MIT

Languages

—

API price (in/out)

No hosted API

Modalities

text

Benchmark results

Bar shows position within the tracked field; marker = field best

Run it locally

VRAM @ Q4

—

VRAM @ FP16

—

Fits on (Q4)

Multi-node cluster required

Throughput data unavailable.

Quantizations

—

Fine-tune it

Permissive

QLoRA84.3 GB1× H200 141GB

LoRA264.1 GB2× H200 141GB

Full fine-tune1990.2 GBbeyond 8× B200

QLoRA SFT on ~10k samples ≈ $64.13 (1× H200 141GB)

Plan a fine-tune →

Ling 3.0 family

Elo progression across releases

Ling-3.0-flashJul 20261000.0 Ling 3.0 TinyAug 20261000.0

Resources

HFinclusionAI/Ling-3.0-flash DOCShttps://developer.ant-ling.com/zh-CN/blogs/ling-3.0-flash-release APIAvailable via inclusionAI API

API price weights · each benchmark row carries its own source badge (see methodology)