Ling-3.2-flash: Ant’s Free 560B Multimodal Model (Specs, Benchmarks, API, Claude Code)

Ant Group’s Ling team has a new flagship. Ling-3.2-flash landed on 5 November 2026, and it is the first Ling model that can look at images and video natively, not just read text.

It keeps the same 560B-parameter body as Ling-3.1-flash, with only 25B parameters active per token, and a 1M-token context window. The headline: Ant’s own tests put it ahead of Claude Opus 5 (max) on three vision benchmarks.

The best part for testers is the price. The API is free for the first two weeks after launch. This guide covers what it is, every number we could verify, how to call it from Python or Claude Code, the limits, and whether you can run it at home.

What Is Ling-3.2-flash?

Ling is the large language model family from Ant Ling, the AI team at Ant Group (the Alipay company). Its open-source side publishes on Hugging Face as inclusionAI.

The “flash” line has moved fast this autumn:

• Ling-3.0-flash: a 124B MoE with open weights, plus a separate vision model, Ling-3.0-flash-VL.
• Ling-3.1-flash: a 560B / 25B-active hybrid-linear MoE aimed at text coding and agents, open weights under MIT.
• Ling-3.2-flash (5 Nov 2026): the same 560B / 25B size, now natively multimodal.

“Native multimodal” means images and video go into the same model that writes code and plans agent steps. Before this, you had to switch to the smaller VL model for anything visual.

Ant lists four focus areas: complex visual understanding, long-horizon multimodal agents, 3D understanding and interaction, and medical multimodal work.

What Ant says it can do

Ant’s launch post and partner write-ups show these demo tasks. They are vendor examples, not independent tests:

• Turning a research PDF with equations and diagrams into working code.
• Turning a screenshot or sketch into a responsive, animated web page.
• Operating Blender through UI actions.
• Converting engineering drawings into 3D web models.
• Recreating 3D game assets from one reference image.

Specs at a Glance

Benchmarks (and How Much to Trust Them)

Ant published three head-to-head numbers against Claude Opus 5 at max effort. All three are vendor-reported release evaluations.

The only third-party signal so far is DesignArena, a crowd-voted arena. Its image-to-frontend board placed Ling-3.2-flash around #10 with an Elo of 1238, as reported by AIbase and AICoder.

Note: Ant picked three benchmarks where it wins. It did not publish text coding scores for 3.2, so for pure coding, Ling-3.1-flash’s numbers (below) are the latest official guide. Artificial Analysis and LMArena had no 3.2 entry when we checked.

For context: Ling-3.1-flash coding numbers

The Ling-3.1-flash model card lists these (vendor-reported): Terminal-Bench 2.1 81.18, SWE-bench Pro 68.74, SWE-bench Multilingual 88.5, HLE 37.44 and SkillsBench 68.7. Ant has not said whether 3.2 keeps, gains or loses on these.

Price and the Free Two Weeks

Ant’s changelog and price page say Ling-3.2-flash is “free for the first 2 weeks after launch”, with original prices resuming after. Counting from 5 November, that is roughly until 19 November 2026; Ant has not posted an exact end time.

The official price page has no Ling-3.2-flash row yet. Some trackers (Sumus, AIbase) quote $0.30 per million input and $0.90 per million output tokens, which matches the level Artificial Analysis listed for Ling-3.1-flash. Treat that as an estimate until Ant publishes it.

Ling-3.1-flash’s list price of ¥2 / ¥6 works out to roughly $0.28 / $0.84 at current exchange rates. If 3.2 lands at the same list price, it would be one of the cheapest frontier-sized multimodal APIs. For a wider view, see our DeepSeek vs Claude price and benchmark comparison.

Warning: Ant’s page tells users to “keep an eye on your account balance”. Billing runs daily at midnight, so a script you leave running after the promo ends will start costing money the next day.

How to Use It: Chat, API and Claude Code

Option 1: Try it in the browser

1. Open the Ling chat. Go to chat.ant-ling.com and sign up or log in.
2. Pick Ling-3.2-flash in the model menu.
3. Upload a file. Try a PDF page, a UI screenshot or a short clip and ask it to explain or rebuild it.

Option 2: Call the API from Python

1. Create a key. In the API console (chat.ant-ling.com/open), create a token. Keys look like sk-studio-....
2. Install the OpenAI SDK. Run pip install openai.
3. Point it at Ling. Use the base URL https://api.ant-ling.com/v1 and model Ling-3.2-flash, as below.

from openai import OpenAI
import os

client = OpenAI(
    base_url="https://api.ant-ling.com/v1",
    api_key=os.environ["LING_API_KEY"],
)

stream = client.chat.completions.create(
    model="Ling-3.2-flash",
    messages=[{"role": "user", "content": "Explain MoE models in 3 bullet points"}],
    stream=True,
)
for chunk in stream:
    if chunk.choices and chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="")

Tip: Use stream=True. Non-streaming calls time out after 90 seconds, and long thinking answers hit that easily. Add extra_body={"thinking": {"type": "disabled"}} when you want fast, short replies.

Other documented options: tools for function calling, temperature (0 to 1), top_p, and enable_search for built-in web search (uses more tokens).

Option 3: Use it inside Claude Code

Ant runs an Anthropic-compatible endpoint, so Claude Code can talk to Ling directly. Ant’s guide shows Ling-3.1-flash; swapping in the 3.2 ID uses the same endpoint.

1. Install Claude Code (see our Claude Agent SDK guide for the SDK side).
2. Set the variables in your shell or a project .env file (keep it out of git).
3. Start Claude Code with the Ling model.

export ANTHROPIC_BASE_URL="https://api.ant-ling.com/anthropic"
export ANTHROPIC_API_KEY="your-ling-key"
export ANTHROPIC_MODEL="Ling-3.2-flash"
export ANTHROPIC_DEFAULT_OPUS_MODEL="Ling-3.2-flash"
export ANTHROPIC_DEFAULT_SONNET_MODEL="Ling-3.2-flash"
export ANTHROPIC_DEFAULT_HAIKU_MODEL="Ling-3.2-flash"

claude -p "Which model are you?"

Mapping all three tiers stops Claude Code from falling back to Anthropic model names. If you switch between many providers, our Claude Code Router guide shows a cleaner way. Ant also lists ready guides for Kilo Code, Hermes Agent and OpenClaw.

Sending Images and Video

Ling-3.2-flash takes image_url and video_url content blocks. Ant’s multimodal tutorial (written for its Ling multimodal models) sets these rules:

import base64

def data_url(path, mime):
    with open(path, "rb") as f:
        return f"data:{mime};base64," + base64.b64encode(f.read()).decode()

resp = client.chat.completions.create(
    model="Ling-3.2-flash",
    messages=[{"role": "user", "content": [
        {"type": "image_url", "image_url": {"url": data_url("ui.png", "image/png")}},
        {"type": "text", "text": "Rebuild this screen as one HTML file with Tailwind."},
    ]}],
    stream=True,
)

For video, swap the block for {"type": "video_url", "video_url": {"url": data_url("clip.mp4", "video/mp4")}}. Keep clips under 30 seconds and trim them first with any editor or ffmpeg.

Ling-3.2-flash vs Ling-3.1-flash vs Rivals

Who should use which

• Screenshot-to-code, PDFs and diagrams: try Ling-3.2-flash while it is free; it is built for this.
• Text-only coding agents you want to self-host: Ling-3.1-flash, because the weights are out under MIT.
• Cheapest vision calls at scale: Ling-3.0-flash-VL is still the budget pick.
• Sensitive or regulated data: stick to a provider your company already approves, or a local model.

Can You Run It Locally?

Not yet. There is no Ling-3.2-flash repo on Hugging Face as of 7 November. Ling-3.1-flash weights followed its API launch, so a release is possible, but Ant has not promised one for 3.2.

Even if weights arrive, this is datacenter hardware territory. Ant’s recommended setup for the same-sized Ling-3.1-flash is 8× H200 (or 4× GB300) with SGLang. A launch sketch using the model card’s image and flags:

docker run --gpus all lmsysorg/sglang:dev-ling-3.1-flash \
  python -m sglang.launch_server --model-path inclusionAI/Ling-3.1-flash \
  --tp 8 --ep 8 --speculative-algorithm NEXTN

Our rough estimate: a 4-bit GGUF of a 560B model needs about 300-330 GB of combined RAM and VRAM, so a 128 GB desktop cannot hold it. For vision on a home GPU, see our best local vision and video LLMs for consumer GPUs.

Limits, Licence and Privacy

• Rate limit: Ant documents a global cap of 2 requests per second; going over returns HTTP 429 “Rate limit reached for requests”.
• Timeout: 90 seconds for non-streaming calls.
• Media: Base64 only, one 30-second video per request, 32 MiB body.
• Licence: the API model has no published weights or licence. inclusionAI says its open models use MIT or Apache licences.
• Privacy: your prompts and files go to Ant’s servers in China. Read Ant Ling’s terms before sending client or personal data.
• Medical use: Ant lists medical multimodal as a focus, but no regulator has cleared it. Do not use it for diagnosis.
• Still unknown: post-promo price, exact promo end time, text coding scores, independent benchmarks and open-weight plans.

Troubleshooting

My Take

Ling-3.2-flash is an easy model to try because it costs nothing for two weeks and drops into tools you already use. If your work is screenshots to code, scanned PDFs or diagrams, run your own test set through it before 19 November.

I would not switch production work yet. Three hand-picked vision scores against Opus 5 are a good sign, not proof, and there are no text coding numbers for 3.2 at all.

The bigger story is Ant’s pace: three flash generations this autumn, with open weights for the first two. If 3.2 weights follow under MIT like 3.1, it becomes one of the most capable open multimodal models around, even if few of us can run it at home.

Frequently Asked Questions

Is Ling-3.2-flash free?

Yes, for the first two weeks after its 5 November 2026 launch, both in chat and over the API. Ant says original prices resume after, but has not posted the 3.2 price yet.

Is Ling-3.2-flash open source?

Not yet. There is no Hugging Face repo for it as of 7 November. The earlier Ling-3.1-flash is open under MIT.

Does Ling-3.2-flash really beat Claude Opus 5?

Only on three vision benchmarks Ant chose, and those are vendor-reported. No independent lab has published a full comparison yet.

What is the context window?

1 million tokens, the same as Ling-3.1-flash.

Can I use Ling-3.2-flash in Claude Code?

Yes. Set ANTHROPIC_BASE_URL to https://api.ant-ling.com/anthropic, use your Ling API key, and map the model variables to Ling-3.2-flash.

How long can videos be?

Up to 30 seconds, one video per request, sampled at 2 frames per second up to 32 frames.

Can I run it on my gaming PC?

No. Weights are not public, and the same-sized Ling-3.1-flash needs around eight datacenter GPUs or roughly 300 GB of memory even at 4-bit.

Who makes Ling?

Ant Ling, the AI team at Ant Group, which publishes open models on Hugging Face as inclusionAI.

Sources

• Ant Ling changelog v2.0.6 (5 Nov 2026): developer.ant-ling.com/en/docs/getting-started/changelog
• Ant Ling API reference (OpenAI-compatible): developer.ant-ling.com/en/docs/api-reference/chat/openai
• Ant Ling Claude Code integration guide: developer.ant-ling.com/en/docs/integrations/claude-code/
• Ant Ling price page, rate limits and multimodal tutorial: developer.ant-ling.com/en/docs/models/price/
• Ant Ling launch post on X (@AntLingAGI, 5 Nov 2026)
• inclusionAI on Hugging Face and the Ling-3.1-flash model card
• AIbase, AICoder, Sumus and the nocode.mba AI release tracker (coverage, DesignArena placement, reported pricing)