Model

Ling-3.1-flash (560B-A25B)

Made by inclusionAI (Ant Group) · License: MIT

A free model for chat and code. It needs an advanced setup.

Will it run on my computer?

Mac

16 GBToo big 24 GBToo big 32 GBToo big 36 GBToo big 48 GBToo big 64 GBToo big 96 GBToo big 128 GBToo big 192 GBTight 256 GBFits 512 GBFits

NVIDIA graphics card

8 GBToo big 12 GBToo big 16 GBToo big 24 GBToo big 32 GBToo big Two 24 GB cardsToo big

AMD Ryzen AI Max

64 GBToo big 128 GBToo big

This is our estimate from the download size. We did not test it on each computer. “Smaller version” means a smaller download than the command below. Tap your computer to see the exact command.

Start it

Install llama.cpp (the steps are in Details).
Paste this command:

llama-server -hf bartowski/Ling-3.1-flash-GGUF:IQ1_S --jinja

What happens next

It downloads the model the first time.
Then open http://localhost:8080 in your browser and chat there.

Sizes and versions

Will it run on my computer?

For the IQ1_S (bartowski) file (131 GB download).

64 GB+ Mac	Won't fit
32 GB graphics card	Won't fit
AMD Ryzen AI Max+ PC, 128 GB (graphics memory set to 96 GB)	Won't fit
128 GB Mac	Won't fit
256 GB Mac	Fits
512 GB Mac	Fits

This is our estimate from published download sizes. We did not test these ourselves.
Fits = download x 1.25 + 2 GB fits in usable memory. That leaves room for longer chats.
Tight = only download + 1.5 GB fits. Short chats work, but long chats or documents may spill onto the CPU and slow down.
Usable memory on a graphics card is all of its graphics memory.
On a Mac it is 2/3 of memory up to 32 GB and 3/4 at 64 GB+. By default macOS keeps part of memory away from the GPU. Apple says a 32 GB Mac gives the GPU 21 GB and a 64 GB Mac 48 GB.
On an AMD Ryzen AI Max+ PC with 128 GB, it is the graphics memory you set in AMD Software. AMD says you can set up to 96 GB.
Apple · AMD

START HERE Advanced: llama.cpp

STEP 1Install llama.cpp (Mac: brew install llama.cpp; Windows/Linux: grab a release from github.com/ggml-org/llama.cpp/releases).

STEP 2Paste the command below. It downloads the model, then serves a chat page at http://localhost:8080.

RUN

llama-server -hf bartowski/Ling-3.1-flash-GGUF:IQ1_S --jinja

SIZE131 GB download (smallest version)

NEEDSA Mac with 192-256 GB of memory (tight on 192 GB) or a multi-GPU server. Use llama.cpp build b11571 or newer.

NOTENo Ollama or LM Studio version as of Oct 19. The smallest file is heavily compressed, so quality drops.

SOURCEsource · checked 2026-10-19

Recommended quants

Official FP8 (~574 GB)
inclusionAI/Ling-3.1-flash
Official recipe: 8x H200 (TP8 + EP8) or 4x GB300 on SGLang.
AtomicChat AD-IQ1_S (141.8 GB) / AD-IQ2_XXS (174.1 GB) / AD-IQ2_M (218.1 GB)
AtomicChat/Ling-3.1-flash-GGUF
Imatrix GGUFs with published KL divergence (IQ2_M: top-1 agreement 92.7% vs BF16). Needs llama.cpp b11571 or newer per the uploader.
AtomicChat AD-IQ4_XS (337.8 GB) / AD-Q4_K_M (354.4 GB)
AtomicChat/Ling-3.1-flash-GGUF
Q4_K_M top-1 agreement 96.9% vs BF16 per uploader.
bartowski GGUFs (IQ1_S 131 GB to Q8_0 608 GB)
bartowski/Ling-3.1-flash-GGUF

Scores and how it compares

Independent score: Artificial Analysis Intelligence Index 41, checked 2026-10-19 · source.

Arena: not on the Arena boards we use (checked 2026-10-19).

Pick this if you have a 192-256 GB Mac (tight on 192 GB) and are comfortable with command-line tools. Artificial Analysis scores it 41.

Skip this if you have a normal PC or Mac. It will not fit, so pick your computer to see the top-ranked model that does.

Compare: Qwen3.8-27B · Pick your computer

BEST PICK ONNot the top pick on any computer page right now.

Settings and apps

Chat template Works

Thinking controlled by chat_template_kwargs enable_thinking (model card recommends true). Reasoning in <think> blocks. Custom architecture (bailing_hybrid), so transformers and vLLM need trust_remote_code.

Fix: On vLLM use --tool-call-parser ling3 --reasoning-parser ling3 and --mamba-cache-mode align (from the official command), or tool calls and thinking won't parse.

Sampling settings

Official	`temp 1.0, top_p 0.95, top_k 20, enable_thinking true`

Runtimes

vLLM	Caveats	Official command uses a vllm branch built from source (git clone -b vllm), not a release wheel. TP8, expert parallel, MTP speculative decoding.
SGLang	Works	Recommended. Day-0 image lmsysorg/sglang:dev-ling-3.1-flash with MTP (NEXTN) speculative decoding.
llama.cpp	Works	Community GGUFs; AtomicChat says b11571 or newer.
Ollama	Unverified	No Ollama library page found as of Oct 19.
LM Studio	Unverified	No listing confirmed as of Oct 19.
MLX	Unverified	No mlx-community quant confirmed as of Oct 19.

Tool calling in agent tools

Claude Code (local endpoint)	Unverified	Model card says it works well in Claude Code, Kilo Code, Qwen Code and Hermes Agent; not tested here.

Sources

MADE BYinclusionAI (Ant Group)

LICENSEMIT · model page · checked 2026-10-19

RELEASED2026-10-13

CHECKED2026-10-19 · 3 sources

Very large open model for agents and coding. Smallest local download is about 131 GB—think big Mac or multi-GPU server.

Known issues

Sources checked

Spotted something wrong or a new fix? Email corrections to [email protected]. Corrections are reviewed before publishing.