# Ling-3.0-flash — Ant Group opens 124B MoE with 5.1B active params under MIT

> A 124B open-weight MoE that runs like a 5B model — Ant Group ships it under MIT.

- Org: Ant Group (inclusionAI)
- Date: 2026-08-05
- Category: model
- Source: https://huggingface.co/inclusionAI/Ling-3.0-flash
- Canonical page: https://ai-tldr.dev/releases/inclusionai-ling-3-0-flash/

## In short

Ling-3.0-flash is Ant Group's 124B open-weight MoE with only 5.1B params active per token. Hybrid Kimi Delta + Gated MLA attention, 56.6 on SWE-Bench Pro, MIT-licensed weights on Hugging Face.

## What it is

Ling-3.0-flash is Ant Group's newest open-weight MoE model, released by their inclusionAI arm. It has 124 billion total parameters but activates only 5.1 billion per token, so it runs closer to a 5B model on hardware while pulling from a much larger expert pool. Weights landed on Hugging Face on August 5, 2026 under an MIT license.

## How it works

The model swaps most standard attention for a hybrid stack — 35 layers of Kimi Delta Attention (a linear attention variant) paired with 7 layers of Gated Multi-head Latent Attention in a 5:1 pattern. The MoE routes each token to 8 of 512 experts plus one shared expert. Ling-3.0-flash was trained in context stages of 8K, then 32K, then 256K tokens.

## Why it matters

The hybrid linear attention combined with SGLang HiCache and Mooncake tiered caching cuts Time-To-First-Token by 60-80% on long inputs, so an agent scanning a large codebase stays snappy. With just 5.1B active params, teams can self-host a strong reasoning model without paying flagship-scale compute costs.

## Who it's for

teams self-hosting a strong coding and reasoning model

## Try it

huggingface.co/inclusionAI/Ling-3.0-flash (SGLang or the inclusionAI vLLM fork; temperature 0.6, top_p 0.95)

## Quick facts

- Maker: Ant Group (inclusionAI)
- Total parameters: 124B
- Active per token: 5.1B
- Architecture: MoE (512 experts, 8 active + 1 shared)
- Attention: Hybrid — 35 KDA + 7 Gated MLA layers (5:1)
- Context window: 256K tokens
- License: MIT

## FAQ

### How is Ling-3.0-flash different from other 100B+ open-weight models?

Ling-3.0-flash mixes linear attention (Kimi Delta) with standard attention in a 5:1 ratio, trained that way from scratch. Ant Group says this cuts Time-To-First-Token by 60-80% on long prompts when paired with SGLang HiCache and Mooncake caching. Combined with its 5.1B active params, it targets long-context agent workloads without flagship compute cost.

### What benchmark scores has Ant Group published for Ling-3.0-flash?

Ant Group's model card reports 56.6 on SWE-Bench Pro, 93.2 on AIME 2026, 87 on HMMT February 2026, and 72.4 on SWE-Bench Multilingual. All figures come from the release page — no independent audit exists yet at open-weights launch.

### How do you run Ling-3.0-flash locally?

Ling-3.0-flash runs on SGLang or a custom vLLM fork published by inclusionAI. Recommended sampling is temperature 0.6, top_p 0.95, top_k 20. The full BF16 checkpoint is 255GB; an FP8 variant at 128GB is also on Hugging Face for smaller GPU setups.

### Was Ling-3.0-flash available before the open-weights release?

Yes, but only through hosted APIs. From July 23 to August 3, 2026 Ling-3.0-flash was free on OpenRouter, Vercel AI Gateway and Kilo. Ant Group open-sourced the weights on August 5 under an MIT license, so anyone can now self-host or fine-tune.

## Links

- [Model Card](https://huggingface.co/inclusionAI/Ling-3.0-flash)
- [FP8 Weights](https://huggingface.co/inclusionAI/Ling-3.0-flash-fp8)
- [Launch Post](https://blog.kilo.ai/p/announcing-ling-30-flash-free-on)
- [Analysis](https://www.digitalapplied.com/blog/ling-3-0-flash-ant-group-efficiency-moe)
