ERNIE Light DT 1.0 vs Qwen 3.8 Flash: Paying More for Less Input

Put ERNIE Light DT 1.0 and Qwen 3.8 Flash on the same line and three numbers do the arguing. Baidu charges $1.13 per million input tokens; A​libaba charges $0.15. Baidu gives you text only; A​libaba accepts text, images and video. Baidu has published no benchmark; A​libaba's model sits behind a documented release with an open-weights twin you can inspect. Qwen 3.8 Flash shipped on 26 August 2026; ERNIE Light DT 1.0 appeared on 10 November 2026, two and a half months later, at seven and a half times the input price.

Newer and dearer is not automatically worse — architectures improve, and a model built around KV-cache compression is solving a problem a general-purpose fast tier does not. But the burden of proof sits firmly on Baidu's side of this comparison, and Baidu has not picked it up.

What Qwen 3.8 Flash actually is

A​libaba's Flash tier is a sparse mixture of experts: 125 billion parameters in the expert body with roughly 6 billion activated per token, plus a 51-billion-parameter n-gram embedding table and about 4 billion of multi-token-prediction heads, which comes to roughly 180 billion parameters on disk. Forty-eight layers, 512 experts, and a hybrid attention stack built from Gated DeltaNet and Q​wen Sparse Attention.

The product SKU is served on QwenCloud with a documented default context of one million tokens, a 991,000-token maximum input and a 131,072-token maximum output. The open-weights artefact with the same architecture is Qwen3.8-Flash-Next, published under the Q​wen Community License, which is a different object with a different licence — useful for inspection, not a substitute for the endpoint.

Notably, Q​wen's sparse attention and Baidu's sparse attention are solving the same problem from opposite directions. Q​wen's QSA operates at the micro-block level rather than selecting individual tokens, which the vendor says cuts long-context latency significantly. Baidu's BSA, paired with feedback-driven history compression, targets cache storage and indexing instead. Both are bets that the future of inference is long context at low cost. They just disagree about which part of the bill to attack.

The comparison, dimension by dimension

• Input price — ERNIE Light DT 1.0 at $1.13 per million tokens vs Qwen 3.8 Flash at $0.15. Baidu is 7.5x dearer.

• Output price — $3.94 against $0.47 per million. Baidu is 8.4x dearer.

• Cached input — $0.113 against $0.0184 per million. Both give a tenfold cached-read discount; the absolute gap is 6.1x.

• Context window — 1,024,000 tokens against 1,000,000, with Q​wen documented at a 991,000-token maximum input. Effectively the same ceiling.

• Maximum output — 131,072 tokens on both. Identical.

• Input modality — text only against text, image and video.

• Reasoning control — four graded levels on Light DT (max, high, low, none) against a flexible thinking mode on Qwen 3.8 Flash that supports a documented maximum-token reasoning budget.

• Tool use and structured output — supported on both, including forced tool choice.

• Independent score — none published for Light DT; Q​wen's Flash-Next sibling has been measured on the Artificial Analysis Intelligence Index, and its score belongs to the open artefact rather than the production SKU.

• Open weights — none for Light DT; a same-architecture checkpoint under the Q​wen Community License for Q​wen.

Where the sevenfold premium could theoretically be earned

It is worth steelmanning Baidu's position rather than dismissing it, because there is a version of this where the price is right.

A model built on compression can, in principle, serve a million-token request with a cache footprint that a conventional model cannot match. If Baidu's BSA really does keep quality flat as context grows — the claim is "maintaining model performance" while cutting KV-cache storage and compute — then the operator's cost per million tokens of context falls. A vendor that has spent research effort on that could reasonably charge more per token while still being cheaper per long request.

The problem is that this is entirely unverified, and the one number that would settle it — effective cache footprint per million tokens — has not been published. There is no model card, no technical report, no benchmark table. Baidu's own product description runs to a single paragraph.

Compare the evidence on the other side. Qwen 3.8 Flash has a vendor page with token-level documentation, an open-weights sibling whose architecture can be read in a config file, and third-party measurements of that sibling. You can disagree with A​libaba's numbers, but you can check them.

Modality is the harder problem

Price is a negotiation. Modality is a wall.

Qwen 3.8 Flash accepts text, images and video, and its product copy names the jobs that follow from that: visual understanding, document and codebase analysis, chart analysis, desktop interaction, long-video analysis. ERNIE Light DT 1.0 takes text and returns text.

If any part of your pipeline involves a screenshot, a scanned page, a chart or a video frame, the comparison is over before the price matters. You either put an OCR or vision model in front of Light DT — adding cost, latency and a failure mode — or you use a model that accepts the artifact. In a support pipeline where half the inbound is a screenshot of an error message, "text only" is not a limitation you work around; it is a different product category.

The counter-case is equally real: retrieval-augmented text work, code analysis on plain source, log triage, long-form document Q&A where the documents are already text. There, the modality edge buys nothing, and the comparison returns to price and quality — where Baidu is still 7.5 times dearer with no evidence to show.

Three ways the premium might actually pay for itself

Rather than assume Baidu is simply overpriced, here are the specific conditions under which choosing Light DT would be defensible — each of them testable in an afternoon.

One: quality holds at 400,000 tokens and above. Ling and Q​wen both advertise a million-token window; whether either degrades in the middle of a very long prompt is exactly what compression schemes claim to fix. If Light DT answers deep-context questions accurately where Qwen 3.8 Flash starts missing the middle of the document, the premium has a case. The test is a set of questions whose answers sit at 5%, 50% and 95% depth in the same long document.

Two: throughput per request is materially better. If compression means fewer tokens of cache per request, a vendor can serve more concurrent long requests on the same hardware — which can translate into lower latency at load. Baidu publishes no throughput figure, and Q​wen publishes no matched one either, so this is measurable only by running both under your own concurrency.

Three: the reasoning ladder earns its keep. Setting effort to none on Light DT gives a fast, cheap path for classification and extraction, while max gives you a thinking model on the same endpoint. Qwen 3.8 Flash has a thinking mode but a less granular ladder. If you are running one endpoint for two very different jobs, that has genuine engineering value — though not obviously $1.13-per-million value.

The realistic read

None of those three conditions is likely on its own to justify a 7.5x input premium, and together they still only beat Q​wen if your workload is long-context, text-only and quality-sensitive at depth. That is a real segment — legal discovery, large codebase Q&A, regulatory document review — but it is not the broad fast-tier market these models nominally share.

The likelier explanation is the least flattering one: Baidu is pricing a new, unmeasured endpoint as a premium product because it can, and the price will move once the market has an opinion. That is a normal pattern for a first listing. It is also why nobody should build a budget on the current number.

Running both without committing to either

Qwen 3.8 Flash is on OrcaRouter's catalogue at A​libaba's list price with no markup applied, which means the price you see on our model page is the price the vendor charges — and if A​libaba cuts it, the cut shows up here the same day rather than at the next contract renewal.

That property matters more than usual in a comparison like this one, because the entire premise is a price spread that nobody has justified and that either party could change. A routing layer that lists both prices side by side, and lets you send five percent of traffic to the new model while keeping ninety-five on the known one, is how you find out whether the premium is real without betting a quarter's budget on it.

Automatic failover is the other half. A model served by a single vendor has a single failure domain; if Baidu has an incident, your long-context path is down. Routing that can fall back to a comparable endpoint when the primary errors is the difference between a degraded feature and an outage.

What to do this week

Start with Qwen 3.8 Flash unless you can name the workload where the extra money buys something. It is cheaper on every unit, accepts more input types, and comes with an inspectable open sibling. At $0.15 and $0.47 per million tokens, a serious evaluation costs pennies.

Test ERNIE Light DT 1.0 only against the specific claim it makes. Build one long-context eval — a 200,000-token document set with answers planted at three depths — and run it on both models. If Light DT holds quality where Qwen 3.8 Flash slips, you have found the case for the premium and can argue it with data. If it does not, you have saved yourself a 7.5x bill on the strength of a vendor paragraph and a name that says "light" while charging like it means "luxury."