Ling-3.2-flash vs ERNIE 5.5 Flash: One Is a Product, the Other Is a Starting Point

Two Chinese labs shipped flagship-adjacent models within hours of each other on 5 November 2026, and almost every comparison you can write about Ling-3.2-flash and ERNIE 5.5 Flash dissolves the moment you notice they are not the same kind of object. Ling-3.2-flash is a finished, instruction-tuned multimodal chat model: text, image and video in, answers out, live on Ant Ling's OpenAI-compatible endpoint and free for a fortnight. What Baidu published is ERNIE-5.5-Flash-Base — the checkpoint at the end of pre-training, before mid-training, before instruction tuning, text-only, and served through a completions endpoint because it has not been taught to follow instructions yet. Baidu's own release page says the post-trained ERNIE 5.5 Flash "will follow."

That single asymmetry decides most of the practical questions below. One model you can call this afternoon. The other you can download, fine-tune, and turn into something — which is a different purchase entirely, with a legal contract attached.

The two objects, side by side

What follows is what each vendor documents today. Where a figure is the vendor's own and unreproduced, it says so.

• Total and active parameters — Ling-3.2-flash 560B total / 25B active; ERNIE-5.5-Flash-Base 280B total / 14B active per token, plus 0.36B of ECHO parameters that both counts exclude.

• Inputs — Ling-3.2-flash takes text, images and video natively, with image_url and video_url blocks in the request; ERNIE-5.5-Flash-Base is text-only, full stop.

• What you receive — Ling-3.2-flash is a post-trained chat model behind an API; ERNIE-5.5-Flash-Base is raw pre-training weights you are expected to post-train yourself.

• Architecture — Ling-3.2-flash is a MoE in the Ling hybrid-linear line Ant has run since 3.1; ERNIE-5.5-Flash-Base is 43 layers (one dense FFN, 42 Multi-Router Latent MoE), 512 routed experts with top-10 routing plus one shared expert, 37 Heavily Compressed Attention layers and six MLA layers with sparse attention.

• Context — Ling-3.2-flash serves a 1M-token window on the API; ERNIE-5.5-Flash-Base was trained at 128K, and its released config enables 1,048,576 tokens for fine-tuning and inference, which is not the same claim.

• Weights — Ling-3.2-flash has none published; ERNIE-5.5-Flash-Base ships FP8 and BF16 safetensors on Hugging Face and ModelScope.

• Price — Ling-3.2-flash is free for two weeks after launch with no published rate for day fifteen; ERNIE-5.5-Flash-Base has no hosted price because Baidu is not serving it, and self-hosting costs whatever 280B of weights costs you.

None of those rows is a quality claim. They are a description of what you would be buying.

What Ling-3.2-flash is actually good for, on the evidence

Ant's launch material is three benchmark numbers and five demo videos. The numbers compare Ling-3.2-flash to Opus 5 at maximum effort on OmniDocBench1.5 (92.21 vs 90.56), ZEROBench(Sub) (43.71 vs 39.52) and Vision2Web (85.86 vs 80.20). All three are vendor-reported, all three are visual, and none has been reproduced by a third party. Read as a set they claim one thing: on document parsing, 3D-scene reasoning and vision-to-web tasks, Ant believes it beats a frontier closed model. They claim nothing at all about text reasoning, coding or agentic terminal work, which is where most production traffic still goes.

The demos are more informative than the numbers, because they show the loop Ant built the model for. Ling-3.2-flash inspects headless renders of material spheres, compares roughness and normal maps, and rewrites its own script; it drives Blender through UI actions; it converts engineering drawings into 3D web models; it reads research PDFs with equations and implements the algorithms in code. Every one of those tasks requires the model to look at an intermediate artefact and correct itself. That is what native multimodal training buys you over a text model with a vision adapter, and it is the capability to test first if this is your use case.

Two operational caveats belong in the same breath. Ant documents a global limit of two queries per second with 429s above it, and a 90-second timeout on synchronous calls with streaming recommended. And the model is free only until roughly 19 November; the pricing page says original rates resume afterwards without saying what they are.

What ERNIE-5.5-Flash-Base is actually good for

Baidu's pitch is cost displacement, and it is a coherent one. The company spent the pre-training compute on 33.5 trillion tokens with the Muon optimizer; you spend yours on domain adaptation, instruction tuning or reinforcement learning, starting from a checkpoint that already has repository-scale code and long-context behaviour baked in. The release is unusually clean as a research baseline: one checkpoint at the end of pre-training, no mid-training, no post-training, so two fine-tuning methods can be compared without arguing about what the base already knew.

Baidu's evaluation compares it to three similarly sized open base models — DeepSeek-V4-Flash-Base (284B-A13B), Hy3-preview-Base (295B-A21B) and MiMo-V2.5-Base (310B-A15B) — under one harness, one prompt set and one judge, across 30 benchmarks in seven categories. That is a better-controlled comparison than most launch tables, and it is still vendor-run. The headline architectural claim worth knowing: 37 of the 43 layers attend over a sequence compressed 128×, and Baidu says the six MLA layers are enough to hold long-context performance. If that holds up under independent testing, it is the most interesting thing in the release — a 128× compression that does not cost you long-context retrieval.

The catch is the licence, and it is not a small one. The ERNIE 5.5 Flash Base Model License Agreement v1.0 defines a Territory that is "all geographical areas worldwide other than the European Union, the United Kingdom and South Korea," and requires that the materials not be used, distributed or displayed outside it. Commercial use is free below US$50 million of annual revenue from related products, provided you send a one-time notice and display "Powered by ERNIE" prominently on a user interface, website, about page or product documentation. Above US$50 million you need Baidu's separate prior written authorisation before you cross the line, and use must be suspended if you cross it without one.

So the honest framing of the "free" 280B checkpoint is: free to download everywhere outside three jurisdictions, free to use commercially up to a revenue threshold, and carrying a branding obligation that most product teams will notice on their marketing site before they notice it in the licence.

Which one to pick, by what you are doing

Pick Ling-3.2-flash if the work is visual and you need an answer this month. Document extraction where layout matters, screenshot-driven UI agents, drawing-to-3D conversion, medical imaging triage, anything where the model has to check its own rendered output. The API is OpenAI-shaped, so the integration is a base-URL change, and the two-week free window means the evaluation costs nothing but your time. Do not put it on a production path yet: there is no published price, no independent score, no weights and no fallback provider.

Pick ERNIE-5.5-Flash-Base if you have post-training capacity and a domain. A 280B-A14B base with a documented architecture, working SGLang and vLLM recipes, an FP8 and a BF16 build, and a Megatron-Bridge fine-tuning example is a genuine asset for a team that wants its own model. Check the Territory clause against your legal entity before you start, not after. And do not benchmark it as a chat model — it continues text, it does not answer requests, so any instruction-following comparison against Ling-3.2-flash is measuring the wrong thing.

If what you actually wanted was a finished ERNIE to call, wait. Baidu says the post-trained ERNIE 5.5 Flash is coming, and until it lands there is no hosted ERNIE 5.5 on Qianfan's price sheet — the international list we read today carries ERNIE 5.1, ERNIE 5.0 and ERNIE 4.5 Turbo, with no 5.5 row.

Neither is on our catalogue — here is what is

Being straight about routing: we re-read the OrcaRouter catalogue on 6 November and it returns 215 model identifiers. There is no Ling, inclusionAI, Ring or ERNIE entry in it; the only "ling" matches are Kling video models. So this comparison is not a choice between two things one key already reaches.

What one key does reach, at provider list price with zero markup, are the models that cover the same two jobs. For vision work with a published rate, Qwen3.8-Flash is $0.15 per million input and $0.47 per million output. For cheap text agents at scale, DeepSeek V4.1 Flash is $0.15 and $0.60, and GLM-5.3-Flash is $0.15 and $0.50. All three take an OpenAI-shaped request, so a Ling-3.2-flash integration ported to them is a model-string change rather than a rewrite — which matters when the free window closes and you still have no rate card.

That is also the argument for putting a router in front of a self-hosted ERNIE fine-tune rather than pointing your app straight at your own cluster: one endpoint, one key, and automatic failover to a hosted model when your GPUs are saturated or your deployment is rolling. The routing DSL composes several models into a single call, which is how you would run a vision model and a text model as one step instead of two.

The thing to watch

Both releases are half-finished in opposite directions, and both gaps close on someone else's schedule. Ant owes a price list, a model card and — if the pattern of 2026 holds — an MIT-licensed checkpoint. Baidu owes the post-trained ERNIE 5.5 Flash, and it owes a clearer statement of whether the Territory clause applies to the model's outputs or only to the weights and their derivatives, which the text as written suggests is broader than most teams will assume.

Until one of those lands, the comparison has a simple shape: Ling-3.2-flash is the one you can evaluate today and cannot budget for, and ERNIE-5.5-Flash-Base is the one you can budget for and cannot call.