Ling-3.0-flash

Updated
21.08.2026
Thinking
Tools
Reasoning
Code
Multilingual
Run

Ling-3.0-flash is a 124B hybrid-linear MoE that activates only 5.1B parameters per token and keeps pace with far larger models on software engineering.

At a glance

  • License: MIT
  • Parameters: 124B total, 5.1B active per token
  • Context length: 262,144 tokens
  • Architecture: hybrid linear attention, 512 routed experts
  • Minimum hardware: about 48 GB of memory at AD-IQ2_XXS

Overview

Ling-3.0-flash is the hybrid reasoning model from inclusionAI, published on August 2, 2026 under MIT. It holds 124 billion parameters and activates 5.1 billion of them per token, which is about a twelfth of what the lab's previous trillion-class flagship activated.

SpecificationLing-3.0-flash
Total parameters124B
Active parameters5.1B per token
ArchitectureHybrid-linear MoE, 512 routed experts plus 1 shared, 8 active
Layers35 KDA blocks and 7 gated MLA blocks
Context window262,144 tokens
ModalitiesText input, text output
ReasoningNative hybrid reasoning, thinking on by default
Release dateAugust 2, 2026
LicenseMIT

The attention stack is where this one differs from most MoE models: five Kimi Delta Attention blocks for every gated MLA block, a layout chosen at the start of pretraining rather than bolted on afterwards. Linear attention blocks carry a fixed-size state, so a long conversation costs far less memory here than on a model that keeps a full KV cache in every layer.

Ling-3.0-flash benchmarks

Vendor-published numbers, each model scored by its own lab, against the previous inclusionAI flagship, the current MoE field and the dense Qwen3.8-27B. A hyphen means the vendor did not publish that benchmark.

Ling-3.0-flashQwen3.8-27BRing-2.6-1TMiniMax-M2.7DeepSeek-V4-Flash
Active parameters
5.1B27.8B dense63B10B13B
AIME 2026
Advanced mathematical reasoning
93.2-95.894.296.5
HMMT Feb 2026
Competition mathematics
87.0-93.571.994.8
SWE-bench Multilingual
Software engineering across languages
72.4-56.776.573.3
SWE-bench Pro
Real-world software engineering, harder split
56.661.753.956.252.6
Humanity's Last Exam
Expert-level questions across every domain
22.730.818.328.134.8

Where the two overlap, Qwen3.8-27B is ahead: 61.7 against 56.6 on SWE-bench Pro, and 30.8 against 22.7 on Humanity's Last Exam. What Ling buys instead is cost per token, 5.1B active against 27.8B dense, and the bigger MoE models keep the lead on mathematics.

Ling-3.0-flash hardware requirements

The system requirement to check is memory. We quantized the model from the original weights and published the builds as AtomicChat/Ling-3.0-flash-GGUF.

MemoryBuild to pickFile size
48 GBAD-IQ2_XXS39.2 GB
64 GBAD-IQ2_M49.1 GB
96 GBAD-IQ4_XXS69.3 GB
128 GB and upAD-Q4_K_S74.2 GB

A 48 GB MacBook Pro runs AD-IQ2_XXS if you raise the GPU memory limit, and a 128 GB Mac Studio has room for the larger builds. On a PC, keep the attention layers on the card and offload the experts to system RAM: with 5.1B active parameters that stays usable on a 24 GB GPU with 64 GB of RAM.

How to run Ling 3.0 Flash in Atomic Chat

Atomic Chat is a free local app for macOS, Windows and Linux. It includes a Hugging Face model browser and a built-in chat, with no manual llama.cpp build required.

  1. Download Atomic Chat for your platform and open it.
  2. Search for Ling-3.0-flash in the model browser and open Download Options.
  3. Pick the build that fits the memory you have, then start a chat.

The full walkthrough, with the quant table and the llama.cpp commands, is in our guide to running Ling 3.0 Flash locally.

If the GGUF ladder is new to you, start with what GGUF is, and compare the format against Apple's in GGUF vs MLX.

License

Ling-3.0-flash is released under the MIT license. It permits commercial use, modification, redistribution and private deployment as long as the copyright notice stays with the code.

Get the weights from Hugging Face

huggingface-cli download inclusionAI/Ling-3.0-flash
# our GGUF build:
huggingface-cli download AtomicChat/Ling-3.0-flash-GGUF --include "AD-IQ2_M/*"
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Ling-3.0-flash",
    "messages": [{"role": "user", "content": "Write a migration plan for this schema."}],
    "temperature": 0.6,
    "top_p": 0.95,
    "top_k": 20
  }'
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "inclusionAI/Ling-3.0-flash"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_id, trust_remote_code=True, torch_dtype="auto", device_map="auto")

messages = [{"role": "user", "content": "Find the off-by-one in this loop."}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
print(tokenizer.decode(model.generate(**inputs, max_new_tokens=2048)[0]))
import OpenAI from "openai";

const client = new OpenAI({ baseURL: "http://localhost:8000/v1", apiKey: "local" });
const res = await client.chat.completions.create({
  model: "Ling-3.0-flash",
  messages: [{ role: "user", content: "Refactor this module into smaller functions." }],
  temperature: 0.6,
});
console.log(res.choices[0].message.content);
Desktop
macOS
(M1 or better)
Download
Windows
(x64)
Download
Linux
(x86_64)
Download

Frequently asked questions

Ling-3.0-flash is a 124B mixture-of-experts model from inclusionAI, released on August 2, 2026 under MIT. It activates 5.1B parameters per token and uses a hybrid linear attention stack with a 262,144-token context window.

Our AD-IQ2_XXS build is 39.2 GB, so 48 GB of unified memory is the practical floor. AD-IQ2_M at 49.1 GB suits a 64 GB machine, and AD-Q4_K_S at 74.2 GB wants 128 GB.

The model alternates five Kimi Delta Attention blocks with one gated MLA block. The linear blocks keep a fixed-size state instead of a KV cache that grows with every token, so long chats cost much less memory than on a standard transformer.

Yes. The weights are MIT-licensed, which allows commercial use, modification and redistribution with no fee.

With offloading, yes. Keep the attention layers on the card and push the experts to system RAM; 64 GB of RAM alongside a 24 GB card runs AD-IQ2_M at a usable speed because only 5.1B parameters are active per token.