ERNIE Light DT 1.0 vs ERNIE 5.5 Flash Base: The Licence Is the Product

ERNIE Light DT 1.0 and ERNIE 5.5 Flash Base are the same lab's answer to two different questions, published five days apart in November 2026. Light DT is an API: no weights, no model card, a $1.13 per million input price and a 131,072-token output cap. Flash Base is a download: 280 billion parameters, 14 billion active per token, 33.5 trillion training tokens, and no hosted endpoint from Baidu at all. The decision between them is not really about benchmarks — neither has a published score worth arguing over — it is about whether you want to rent a serving stack or own a licence to build on one.

And the licence is where this comparison stops being symmetrical, because Flash Base's is unusually specific about where you may use it and who you must tell.

Two artefacts, one lab, five days apart

Flash Base landed on 5 November 2026 with a technical report and open weights on Hugging Face in two precisions — FP8 and BF16 — released under the ERNIE 5.5 Flash Base Model License v1.0. It is a pre-training checkpoint, meaning the weights you get are the ones that came out of pre-training, before any instruction tuning or preference work. There is no chat template worth the name and no instruct sibling published alongside it.

Light DT arrived on 10 November 2026 as an international API endpoint, described by Baidu as based on open-source G​LM models with research from the company's Search "Lightning" line folded in, and built around Baidu Sparse Attention plus a feedback-driven history compression scheme that the vendor says cuts KV-cache storage and compute. It is post-trained, it answers questions, it calls tools, and it does not exist as a file you can download.

Put that way, the pairing looks lopsided in the obvious direction: one is a finished product and the other is a starting point. But the price says otherwise, and the licence says something else again.

The rate card that runs the wrong way

Light DT costs more per token than a model with twenty times its active parameter count. That is the fact to sit with before anything else in this piece:

• Input price — ERNIE Light DT 1.0 at $1.13 per million tokens vs ERNIE 5.5 Flash Base at $0, because there is no Baidu endpoint and therefore no list price. The $0 is real: hosting it is a capital cost, not a line on an invoice.

• Output price — $3.94 per million for Light DT; for Flash Base, your own GPU-hours.

• Cached input — $0.113 per million on Light DT, a tenfold discount on cached reads. Flash Base has no equivalent because cache economics become your scheduler's problem.

• Active parameters per token — undisclosed for Light DT; 14 billion for Flash Base.

• Context — 1,024,000 tokens for Light DT; 1,000,000 for Flash Base in fine-tuning and inference, with 128,000 as the maximum training context.

• Maximum output — 131,072 tokens for Light DT; bounded by your serving configuration for Flash Base.

• Weights — none for Light DT; 280B in FP8 or BF16, split across 128 safetensors shards in the FP8 repository, for Flash Base.

• Serving stack — Baidu's, for Light DT; SGLang or vLLM with Baidu's own plugin, for Flash Base. The published SGLang recipe asks for tensor parallelism of 4 and an attention data-parallel size of 4, which is a four-GPU floor before you have loaded a single request.

What self-hosting 280B actually costs per million tokens

This is the calculation that decides the matchup, and it is worth doing with round numbers rather than vibes.

Four GPUs is the floor for the published configuration, but 280B parameters in FP8 is roughly 280 GB of weights before any cache, so a realistic production node is eight accelerators with room for KV cache and batching. Call it eight H100-class cards at a blended rental of $2 per card-hour — a mid-market 2026 figure, not a quote — and the node costs about $16 an hour, or roughly $11,700 a month if it runs continuously.

Now the throughput question. A sparse MoE firing 14 billion parameters per token is fast per token in the way that all sparse models are fast, but aggregate throughput on a batch is what sets your unit cost, and Baidu publishes no tokens-per-second figure for this checkpoint. Suppose, conservatively, that the node sustains 4,000 output tokens per second across a full batch. That is 14.4 million output tokens an hour, which at $16 an hour lands near $1.11 per million output tokens — before utilisation losses, before the engineer who owns the deployment, before the idle hours when traffic dips.

Run it at a third of capacity, which is what most real deployments do outside peak, and the effective cost triples. The honest summary is that self-hosting Flash Base costs somewhere between roughly one and three dollars per million output tokens depending on how well you fill the node — a range that straddles Light DT's $3.94 and can beat it comfortably at high utilisation.

Which means the price premium on Light DT is not a tax on ignorance. It is what Baidu charges for not having to do that calculation, and for a compression scheme it claims keeps the cache small enough that a million-token window stays affordable to serve. If your traffic is spiky, low-volume, or mostly long-context, paying $3.94 and letting someone else own the utilisation problem is the rational choice.

The licence, which is where Flash Base gets complicated

Flash Base's licence is not a standard permissive one, and three clauses matter more than any benchmark.

• Territory. The grant applies to "all geographical areas worldwide other than the European Union, the United Kingdom and South Korea." Using the weights, or distributing derivatives, outside that territory falls outside the rights granted. If you have EU users, an EU inference region, or EU staff touching the model, this is a blocker to read carefully with counsel rather than a footnote.

• Revenue threshold. Commercial use is free below $50 million in annual revenue related to the products or services using the model, subject to a one-time notice to Baidu. Above that, you need separate written commercial authorisation before you start, and if you grow through the threshold you must get authorisation before crossing it or suspend affected use.

• Attribution. Commercial products must display the exact words "Powered by ERNIE" in a readily accessible place — a UI, an about page, or, for APIs and backend services, the developer documentation.

There is also a clause worth flagging for anyone planning to distil: the licence forbids using the materials or their outputs to improve any other AI model other than the model itself or its derivatives. Training a smaller student on Flash Base outputs is inside the grant; training a different architecture on them is not.

Light DT has none of these problems because it has no weights to license. You get a terms-of-service agreement and a rate card, which is a much smaller surface area for a legal team and a much smaller surface area for ambition.

Capability: what each one will and will not do out of the box

• Reasoning control — Light DT exposes four effort levels (max, high, low, none) with max as the default. Flash Base has no reasoning mode at all; it is a completion model, and any reasoning behaviour is something you would have to train in.

• Tool calling — Light DT supports function calling with forced tool choice, structured outputs and JSON response formats. Flash Base supports whatever you implement around it.

• Modality — both are text in, text out. Neither takes images, and neither will read a scanned document for you.

• Long context — Light DT advertises 1,024,000 tokens through its compressed-cache design. Flash Base supports 1M during fine-tuning and inference, but at that length on eight cards your KV cache budget is the binding constraint, and the compression Baidu built into Flash's own attention — Heavily Compressed Attention on 37 of its 43 layers — is what makes it feasible at all.

• Evaluation — Flash Base publishes a full vendor table: MMLU 93.3, MMLU-Pro 78.4, GPQA 60.2, HumanEval+ 82.3, LiveCodeBench 40.1, HELMET-RAG 71.5, LongBench-v2 48.7, measured against DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base under one pipeline. Those are Baidu's own numbers, run by Baidu, and should be read as such. Light DT publishes nothing at all.

That last line is the asymmetry that matters most for a buyer. Flash Base at least lets you check its claims — download the weights, run your own eval, and see whether the vendor table holds on your data. Light DT asks you to take the mechanism on faith, because the only evidence available is a paragraph of product copy.

Where a router fits, and where it does not

Neither of these models is on OrcaRouter's catalogue today, and saying otherwise would be the kind of thing that costs a reader a week. What a router does change is the shape of the decision.

If you take the Flash Base route, you are committing capital and people to a model whose quality you can measure but whose serving cost you cannot fix — and the sensible architecture is to keep a hosted model behind the same interface as a fallback for the days your node is full, your batch is cold, or an evaluation run is hogging the cards. If you take the Light DT route, you are committing to a single vendor's endpoint with no second provider behind it and no weights to move if the price changes.

Either way the failure mode is the same: a long-context workload pinned to one model with no way to compare what the alternative would have cost. That is what a routed endpoint with a per-model price list solves — not by making the choice for you, but by making the cost of the choice visible per request instead of per quarter.

Which one to pick

Take ERNIE 5.5 Flash Base if you have a training problem rather than a serving problem: you want a 280B sparse starting point, you intend to fine-tune it, your users and infrastructure sit outside the EU, the UK and South Korea, and you are under the revenue threshold or prepared to ask Baidu for authorisation. The vendor benchmark table gives you a defensible baseline, and the weights mean nobody can take the model away from you.

Take ERNIE Light DT 1.0 if the context window is the point. A million tokens served through a compressed cache, with a tunable reasoning budget and tool calling, is a specific product for a specific job — long retrieved documents, long agent transcripts, long-running sessions — and paying a premium per token to avoid owning the cache budget is a legitimate trade.

Take neither if your prompts are short. At 8,000 tokens a request, the entire architecture story on both sides is irrelevant, and $1.13 per million input tokens is simply an expensive way to do what a model at a tenth the price will do just as well.