ERNIE 5.5 Flash vs ERNIE 5.5 Flash Base: One You Can Download, One You Can Only Rent

Baidu has now shipped both halves of the same model, in the wrong order. ERNIE 5.5 Flash Base — the 280-billion-parameter checkpoint, 14 billion active per token — went up on Hugging Face on 5 November 2026 with a twenty-page technical release, thirty benchmark rows and a licence that took a lawyer to read. Six days later, on 10 November, ERNIE 5.5 Flash appeared as a priced endpoint on Baidu Qianfan at $0.28 per million input tokens and $1.13 per million output, with no announcement, no model card and no evaluation table whatsoever. So the choice between ERNIE 5.5 Flash and ERNIE 5.5 Flash Base is not a choice between a big model and a small one. It is a choice between the one artefact you can inspect and the one you can actually call. The baselines you would benchmark either ERNIE against are billed at provider list prices, with the OrcaRouter fee shown as $0.00 on every row so a comparison run costs one key and one invoice.

Same weights, two different finished states

The base checkpoint is the end of pre-training and nothing else. It completes text. It does not follow instructions, hold a system prompt, call a tool or decline a request, because those behaviours are produced by post-training and this checkpoint has had none. Baidu says so directly: the repository is "the checkpoint at the end of pre-training, before any mid- or post-training," positioned as a starting point for domain adaptation, instruction tuning and reinforcement-learning research.

ERNIE 5.5 Flash is what post-training made of it. It answers prompts, exposes function calling and structured outputs, and reasons before it speaks. The two are not versions of each other in any usable sense — one is a training input, the other is a product — and the practical difference is the entire decision.

• Parameters — 280B total, 14B active per token, sparse mixture-of-experts, on both sides. Same architecture.

• Download size — 288 GB across 139 files in the FP8 repository, with a separate BF16 repo alongside it. This is the figure to budget for, not the parameter count doubled.

• Context — 128,000 tokens at training time, with the released configuration set to 1,048,576 for fine-tuning and inference. The served model is listed at 1,024,000 input, which is the same window rounded down for a rate card rather than a different capability.

• Output ceiling — the base checkpoint has none to speak of; the served model allows 921,600 output tokens because it reasons on every request and cannot be told not to.

• Modality — text in, text out, both sides. No images, no audio, no PDFs. Any workload with a screenshot in it is out of scope for both, and that is not going to change with a fine-tune.

• Price — the weights are free under a licence with territorial limits; the API is $0.28 and $1.13 per million tokens, with cached input at $0.028.

• Independent score — neither has one. Artificial Analysis carries ERNIE 4.5 and ERNIE 5.0 rows and no ERNIE 5.5 row of any kind.

The documentation runs exactly backwards

For a base checkpoint, Baidu published a genuinely detailed technical release: the architecture (43 layers, 512 routed experts per MoE layer with top-10 selection plus a shared expert), the training budget (33.5 trillion tokens with the Muon optimizer), the attention design, the tokenizer, the serving recipes for vLLM and SGLang, and a benchmark table that includes rows where the model loses. ChineseSimpleQA reads 73.0 against D​eepSeek's 75.3. BBH with chain-of-thought is a three-way tie inside noise. That is the shape of a table somebody built rather than a table somebody picked, and it is the kind of disclosure that earns a model a fair hearing.

For the chat model, Baidu published a price. No knowledge cutoff. No evaluation. No statement about how the post-trained model differs from the checkpoint in anything but behaviour. The description attached to the endpoint is three sentences about architecture and speed — 280 billion total parameters, 14 billion active, a one-million-token window, more than 200 tokens per second — and says nothing about quality.

The consequence is that the two artefacts are being judged on incompatible evidence. Everything knowable about ERNIE 5.5 Flash Base comes from Baidu's own harness against baselines Baidu chose, and all of it describes a model that cannot follow an instruction. Everything knowable about ERNIE 5.5 Flash is a rate card. Anyone who quotes the base checkpoint's MMLU-Pro of 78.4 as evidence about the chat model is making the single most common error available in this comparison: base-model evaluations and instruction-model evaluations measure different things and cannot be subtracted, inherited or extrapolated.

The licence is the part people skip

The weights are downloadable, which is not the same as open source. The ERNIE 5.5 Flash Base Model License Agreement v1.0, dated 5 November 2026 and governed by Hong Kong law, restricts the grant by territory and by revenue:

• Territory — use, reproduction, modification and distribution are licensed everywhere "other than the European Union, the United Kingdom and South Korea." For a company incorporated in any of those three, this is not a commercial-terms problem, it is a no.

• Revenue — free up to US$50 million in annual related revenue, and a one-time notice to Baidu is required even below that line. Above it, use must pause until written authorisation arrives. Related revenue is aggregated across products and allied companies, so the ceiling arrives earlier than a single product's P&L implies.

• Attribution — every commercial product must display "Powered by ERNIE" prominently.

• Outputs — may not be used to improve any other model. Training a derivative is permitted; distilling this one into something else is not.

None of that applies to calls made through Baidu's API. The hosted route is governed by commercial API terms, not by the weights licence, so a European team excluded from the download can still rent the model by the token — which is, quietly, one of the stronger arguments for the endpoint over the repository.

Which one you actually want

The decision falls out of three questions, and none of them is "which is smarter."

Can you run 288 GB? If not, the base checkpoint is not an option regardless of its terms, and the comparison collapses to "use the API or use a different model." A 280B mixture-of-experts with 14B active is cheap to serve per token and expensive to host per machine, and the gap between those two facts is where most self-hosting plans die.

Is your task fixed and high-volume? This is the only case where the weights win on economics. A single narrow task — classification, extraction, a domain-specific transform — fine-tuned onto a base checkpoint amortises the training run and then costs nothing per call. Paying $1.13 per million output tokens forever for a task you could own is a slow leak, and the base checkpoint is priced, in licence terms, precisely for that buyer.

Do you need it to work this week? Then it is the endpoint, because a fine-tune on 280B parameters is a budget line and a calendar quarter, not an afternoon. The API gives you a tool-calling, reasoning, million-token-input model today, undocumented, at a price that is roughly double the Flash-tier competition and not obviously wrong for it.

What no one should do is treat the two as interchangeable in an evaluation. Base-checkpoint scores are not a preview of API quality, and the API's price is not evidence of API quality. The only measurement that settles anything is your own — a fixed prompt set run through the endpoint, counted in output tokens, compared against whatever you were going to use instead.

Running both questions through one key

The comparison half of this work does not require a Baidu account. OrcaRouter routes more than 200 models behind one OpenAI-compatible key at provider list price with no markup: MiniMax M3.1 Flash at $0.15 and $0.60, GLM-5.3-Flash, DeepSeek V4.1 Flash, Qwen 3.8 Flash and Gemini 3.8 Flash are all there, and automatic failover keeps a benchmark run alive when a provider wobbles mid-sweep. ERNIE 5.5 Flash is not among them — there is no ERNIE entry in the catalogue and no Baidu endpoint for us to route to — so the endpoint has to be called on Baidu's own API and its numbers carried across by hand.

That asymmetry is worth stating plainly rather than working around: the model you can download is free and unusable without a training run, the model you can call is undocumented and reachable through exactly one vendor, and the models you would benchmark either against are on a single routed key. Two of those three facts will change. The licence clause probably will not.

What to watch is narrower than the model itself. A model card for ERNIE 5.5 Flash with an evaluation table would make this comparison answerable. A second provider serving the preview would make its uptime and its price competitive. And an independent score from a tracker that runs its own harness would settle in an afternoon what a fortnight of specification-reading cannot — because right now the honest summary is that Baidu has documented the half of the model nobody can use and left the half everyone can use to be taken on trust.