ERNIE Light DT 1.0 Costs Four Times What Baidu's 280B Flash Model Does

Baidu's newest small model is priced like a large one. ERNIE Light DT 1.0, listed on 10 November 2026, carries a list price of $1.13 per million input tokens and $3.94 per million output tokens. The ERNIE 5.5 Flash Preview — 280 billion parameters from the same lab, a bigger output ceiling, a far smaller rate card — is listed at $0.28 and $1.13. Read those two lines together and the name stops making sense: the model called Light is the expensive one, by a factor of four on input and three and a half on output.

That inversion is not a typo in a rate card, and it is the first thing worth understanding about this release. Baidu has not shipped a cheaper Flash. It has shipped a different kind of object — one built for a cost problem that does not show up on a per-token price list at all, and priced as if the thing it saves you is worth more than the tokens it charges for.

What was actually published, and when

The model appeared as baidu/ernie-light-dt-1.0-intl — the -intl suffix marking the international endpoint rather than the mainland Qianfan SKU — with a snapshot stamp of 10 November 2026. It is served by Baidu directly, not through a reseller network, and it is text in, text out. No images, no audio, no files.

The two numbers that define its shape are a 1,024,000-token context window and a 131,072-token maximum output. Those are not unusual in late 2026; a million-token window is table stakes in the Chinese open-weight tier and has been since the summer. What is unusual is a million tokens on a model whose own vendor copy describes it as light.

Baidu's description is short, and worth quoting rather than paraphrasing, because it is the only technical statement the company has made: the model is "based on open-source G​LM models" and "includes research findings from Baidu Search's Lightning large model," and it "features Baidu Sparse Attention (BSA), a unique sparse attention mechanism developed by Baidu ERNIE."

Sparse attention, history compression, and why a Light model gets a million tokens

Three claims sit inside that paragraph, and they are doing different work.

The first is lineage. A Baidu model built on G​LM is not a small detail — G​LM is Z.ai's architecture family, and Baidu saying so out loud tells you this is an adaptation strategy rather than a from-scratch pre-training run. Baidu is not claiming a new foundation model here. It is claiming a better way to run one.

The second is the Search connection. Baidu Search's "Lightning" line is the model family behind the search product's own answer generation, which means it has been optimised under a very specific pressure: enormous query volume, long retrieved contexts, and a hard latency budget. Research findings that survive that environment tend to be about memory traffic, not about benchmark scores.

The third — and the one that explains the price — is BSA plus what Baidu calls "feedback-driven history compression," which it says "improves the storage and indexing of historical information" and "significantly reduces KV cache storage and computational costs while maintaining model performance."

Strip the marketing and that is a KV-cache story. In a standard transformer, the cost of holding a long conversation or a long document in context grows with every token you keep, because every layer caches keys and values for every position. That cache is what makes a million-token window expensive to serve even when the model itself is small. Compression schemes attack exactly that: keep a compact, indexed representation of the distant past instead of the raw per-token cache, and the long-context bill falls.

That is the same problem Baidu solved differently in ERNIE 5.5 Flash, where 37 of 43 layers use Heavily Compressed Attention — caching one entry per 128 tokens — while six layers keep full token-level access through MLA with sparse attention. Light DT appears to push further in the same direction: more compression, better indexing of what was compressed. Baidu has not published layer counts, parameter counts, or the compression ratio, so how far it pushes is unknown.

What follows from the architecture is a pricing hypothesis, and it is only a hypothesis because Baidu has not said so: if your workload is dominated by cache residency rather than by fresh tokens, the vendor can charge more per token and still be the cheaper option per request. Whether that is true depends entirely on a number Baidu has not published — the effective cache footprint per million tokens of context.

The price, in the units a buyer actually uses

Per-token rates, as listed:

• Input — ERNIE Light DT 1.0 at $1.13 per million tokens against ERNIE 5.5 Flash Preview at $0.28. Light DT is 4.0x the price.

• Output — $3.94 against $1.13 per million. Light DT is 3.5x the price.

• Cached input — $0.113 against $0.028 per million. Light DT is 4.0x the price, and both discount cached reads by a factor of ten against fresh input.

• Context — 1,024,000 tokens for both.

• Maximum output — 131,072 tokens for Light DT against 921,600 for ERNIE 5.5 Flash Preview. This is the one line where the small model is genuinely smaller, by a factor of seven.

• Modality — text only for Light DT; text only for the Preview endpoint as well.

The output ceiling is the tell. ERNIE 5.5 Flash Preview has been built to emit very long answers — nine hundred thousand tokens of output is not a chat model's number, it is a number for batch generation, long-form synthesis, or agent transcripts that never get trimmed. Light DT's 131,072 is a normal, generous, unremarkable output budget. Baidu has deliberately capped the thing that costs the most to serve.

So the shape of the product is: charge a premium on every token that enters, and refuse to let the model ramble. That is coherent pricing for a model whose expensive resource is the context it holds, not the text it writes.

What the endpoint will and will not do

The capability surface is wider than the price suggests, and narrower than the name implies.

• Reasoning is optional and tunable. Four effort levels are exposed — max, high, low and none — with max as the default. Being able to set effort to none matters for a model in this tier: it is the difference between a latency-sensitive classifier and a thinking model, on the same endpoint.

• Tool calling is supported, including forced tool choice. Function calling, structured outputs and JSON response formats are all available, which makes it usable as the planner inside an agent loop rather than only as a text generator.

• Sampling is conventional — temperature, top_p, max_tokens — with nothing unusual exposed and no sign of the exotic decoding knobs some Chinese labs ship.

• It takes text and returns text. If your pipeline feeds it a PDF scan, a screenshot or a call recording, you need an OCR or transcription step in front of it, and that step has its own cost and its own failure modes.

• It is served by one provider. A single-vendor endpoint has no failover behind it, so an incident at Baidu is an incident in your product. That is a procurement fact, not a criticism.

What is missing

Three gaps, and they are all in the same place: the evidence.

There is no benchmark. Not a vendor table, not a chart, not a single number — Baidu's release copy makes a mechanism claim and stops. No independent evaluator has a row for it either. Artificial Analysis, which carries rows for the ERNIE 5.5 family and for most of the Chinese fast tier, does not score this model. So the honest statement of its quality today is: unknown, on every axis, including the axis Baidu is implicitly selling, which is whether compression costs you accuracy at long context.

There are no weights. ERNIE 5.5 Flash Base shipped open weights on 5 November 2026, 280B parameters with a licence that carves out the EU, the UK and South Korea. Light DT ships as an API only, with no public repository and no model card. If you want to inspect the compression scheme, you cannot.

And there is no definition of "DT." Baidu has not published a glossary entry, and the two letters do not map onto anything in the company's existing naming — where Flash means a cheap tier, Base means a pre-training checkpoint, and Preview means a hosted instruct model. Until Baidu says what DT stands for, treat it as a product code, not a description.

How to call it, and what to do in the meantime

For now, ERNIE Light DT 1.0 is reachable through Baidu's own API and through third-party platforms that carry the international endpoint. It is not on OrcaRouter's catalogue — check it against the live model list before you plan around it — so this is not a piece where we can hand you a key and a price and be done.

What is worth doing today is measuring the thing the price depends on. Take the longest context your product actually sends — the retrieved documents, the rolling conversation, the codebase slice — and send it to a model you can already call. Record the input tokens, the cached tokens if the provider reports them, and the wall-clock latency at the deep end of the window. That number, not the per-token rate, is what decides whether a compressed-cache model is a saving or a premium. If your traffic never exceeds 32,000 tokens, the architecture Baidu is selling buys you nothing, and $1.13 per million is simply four times too much.

The second thing worth doing is keeping the swap cheap. The reason a model with no benchmarks and a fourfold premium is even discussible is that the alternative to trying it is not a rewrite — it is a base URL. Route your long-context traffic through one endpoint with the per-model price exposed, and a model like this can be put behind ten percent of traffic on a Tuesday and removed on a Wednesday. That is the only responsible way to evaluate an unmeasured release: small exposure, real traffic, and a price you can read per request rather than per invoice.

What to watch, in order: a model card with parameter counts and a compression ratio; a first independent evaluation, ideally one that reports at 128K and 1M separately rather than a single averaged score; and a price move. That last one is the most informative. A fourfold premium on a light model is either a statement about cache economics or a placeholder, and the way to tell the difference is whether it survives its first quarter.