ERNIE-5.5-Flash-Base: Baidu Opens a 280B Checkpoint, Then Narrows Where You May Use It

For two weeks, the most notable thing about ERNIE-5.5-Flash-Base was how little Baidu said about it. The Hugging Face repository was registered on October 21, 2026 and sat essentially empty until November 2, holding little more than a README. The first weight shards landed that morning, and the 128-part FP8 checkpoint finished uploading the next day. Only on November 5 did Baidu publish the technical blog and the model card that explain what it had actually put there: a 280B-parameter sparse mixture-of-experts checkpoint, 14B of them active per token, pretrained from scratch on 33.5 trillion tokens.

By that morning the primary repository had been downloaded four times, against a like count in the high fifties. The distance between those two numbers is the release in miniature. A great many people are watching Baidu's second turn toward open weights, and almost nobody has yet taken the weights home.

Two things explain the hesitation, and both are properties of the release itself rather than chatter around it. The first is that this is a base checkpoint — the end of pre-training, with no mid-training and no post-training — so it is not a chat model and does not behave like one. The second is the licence. It is not Apache 2.0, it does not extend to the European Union, the United Kingdom or South Korea, and its commercial terms change once your related revenue passes US$50 million a year.

What Baidu published, and in what order

The commit history on the repository is a cleaner account of the release than any announcement. An initial commit on October 21. A README on October 26, followed two minutes later by a folder upload and then a deletion of the architecture and performance images — the ordinary churn of someone assembling a page. The configuration files arrived on November 2 at 06:29, the weight shards immediately after, and the last of the 128 before 09:00 the following morning. The benchmark and architecture figures went up on November 4. The logo, the tokenizer tweaks and the final README passes came on November 5, within hours of the blog post going live.

So the sequence is: weights first, documentation second, announcement third. Baidu did not stage a launch and then drag the files out behind it. It put a 280B checkpoint on the internet and wrote the explanation afterwards. That is worth stating plainly, because the framing you will see in the Chinese-language coverage — a triumphant open-source announcement — describes what happened on November 5, not what happened on November 2.

Two repositories exist, and the split matters more than it usually does. baidu/ERNIE-5.5-Flash-Base ships FP8 mixed-precision safetensors; baidu/ERNIE-5.5-Flash-Base-BF16 ships the same checkpoint in BF16. Both are ungated, both are text-only, and both declare the same 280B total and 14B activated. A third set of weights exists inside the model as ECHO parameters — 0.36B of them — which Baidu explicitly excludes from the headline counts.

The counting conventions are worth pinning down, because they disagree on the same page. Baidu's 280B includes the input embedding and the output head, and its 14B activated figure counts the output head but not the embedding lookup. Hugging Face's own model-size widget reads 286B — a different convention applied to the same files. Neither is wrong; they are answering slightly different questions, and a spec sheet that mixes them will not add up.

A base checkpoint is not a chatbot, and the difference is not cosmetic

The single most consequential line on the model card is the least dramatic one: this is "the checkpoint at the end of pre-training, before any mid- or post-training." Baidu's stated purpose is domain adaptation, instruction tuning and reinforcement-learning research. It is a starting point, sold as a starting point.

What that means in practice is that you cannot swap it into a chat product and expect ERNIE. The card serves it through the /v1/completions endpoint rather than chat completions, and it notes that inputs need a BOS token, so tokenization has to keep add_special_tokens=True. A base model continues text. It does not follow instructions, refuse requests, or hold a conversation, because nothing has taught it to.

There is also no post-trained sibling yet, and this is the part readers most often get wrong when they see the benchmark numbers. Baidu's blog says the post-trained ERNIE 5.5 Flash "will follow." It does not exist today. There is no hosted API for it on Qianfan that the release materials name, no chat model to compare against, and no row for ERNIE 5.5 Flash on Artificial Analysis — I checked the models index on November 5 and it carries ERNIE 4.5 300B A47B and ERNIE 5.0 Thinking Preview, and nothing from the 5.5 line.

So the honest description of 5.5 Flash as a product family, on the day of its announcement, is: one base checkpoint, open weights, no instruct model, no endpoint. Everything else is a promise about later.

Inside the architecture, briefly

Baidu spent its engineering effort on the cost of long context, and the design is legible even if you never train it.

• Attention — 37 of 43 layers use Heavily Compressed Attention, which caches one entry per 128 tokens, while 6 layers keep token-level access through MLA with sparse attention (DSA). At a matched KV-cache budget, Baidu reports HCA paired with MLA beats HCA paired with compressed sparse attention.

• Head geometry — the cheap, 128×-compressed HCA layers get wide 512-dimension heads; the expensive token-level MLA layers get narrow 256-dimension ones. Head width is matched to what each layer's attention actually costs.

• Experts — 512 routed experts per MoE layer with top-10 routing plus one shared expert, dispatched and combined in a 2,048-dimension latent space at half the hidden width. Load balancing is auxiliary-loss-free.

• Residuals and output — Manifold-Constrained Hyper-Connections with four parallel residual streams replace the standard residual path, and a MultiMax head with a learnable piece-wise temperature sharpens the vocabulary distribution.

• ECHO — instead of a separate multi-token-prediction module, the backbone's own final layer predicts future tokens and doubles as a speculative-decoding draft model. This is the 0.36B that sits outside the parameter count.

• Tokenizer — roughly 201,000 BPE entries, with numbers tokenized right-to-left in groups of up to three digits.

The context story is the practical headline: 128K during training, and 1M for fine-tuning and inference, with the released configuration already enabling the 1M setting. That is a genuinely large window for an open checkpoint at this size, and it is the capability Baidu built the architecture around rather than bolted on.

The benchmark table is real, and it is not the number you think it is

Baidu publishes 30 rows across exam, knowledge, STEM, reasoning, multilingual, code and long context, against three base-model contemporaries: DeepSeek-V4-Flash-Base at 284B total and 13B active, Hy3-preview-Base at 295B and 21B, and MiMo-V2.5-Base at 310B and 15B. All four were run through one evaluation pipeline with the same prompts and few-shot settings — which is the right way to build the table, and also the reason it cannot be dropped into a leaderboard next to chat models.

ERNIE-5.5-Flash-Base leads 24 of the 30 rows. The margins that matter most are in code and exam-style knowledge, where base-model differences are usually small:

• HumanEval+ (1-shot, Pass@1) — ERNIE 82.3 vs DeepSeek 69.5, Hy3 78.0, MiMo 74.4

• MMLU-Pro (5-shot, EM) — ERNIE 78.4 vs DeepSeek 69.2, Hy3 66.9, MiMo 65.7

• LiveCodeBench (1-shot, Pass@1) — ERNIE 40.1 vs DeepSeek 31.9, Hy3 33.5, MiMo 32.4

• RepoQA (0-shot, Pass@1) — ERNIE 91.7 vs DeepSeek 81.0, Hy3 91.3, MiMo 56.5

It also loses six, and those six are more informative than the wins. ChineseSimpleQA goes to DeepSeek at 75.3 against ERNIE's 73.0. BBH with chain-of-thought goes to Hy3 at 90.0 against 89.2. INCLUDE-lite is a rounding-distance loss to DeepSeek, 79.8 to 79.5. FullStackBench goes to Hy3 at 56.7 against 54.6. HELMET-Rerank is the widest defeat, 60.0 to 56.2. InfiniteBench, 48.0 to 46.5, also goes to DeepSeek — which is a pointed result for a model whose central design goal was long context.

Every one of those figures is Baidu's own, produced by Baidu, on a harness Baidu configured, and published alongside the weights. No third party has reproduced any of it. That is not an accusation; it is the normal state of a model that is hours old, and it is exactly the state in which vendor tables are most often quoted as if they were measurements.

The comparison that would actually be useful does not exist yet, and it is not the one the table invites. These are all base models. A 93.3 on MMLU for a base checkpoint is not a 93.3 on a chat model, and putting it beside an Artificial Analysis Intelligence Index score — which is computed on post-trained models doing agentic and reasoning work — would be a category error. Baidu has published a careful within-category comparison. The mistake available to everyone reading it is to treat it as a cross-category one.

The licence is the part that will actually stop people

Strip the announcement away and the document that governs this release is the ERNIE 5.5 Flash Base Model License Agreement v1.0. It is dated November 5, 2026 — the licence itself is the most reliable confirmation of the announcement date — and it is a custom agreement, not an OSI-approved one.

Three clauses do most of the work.

• Territory. The grant is made "solely within the Territory," which Section 1.11 defines as every geographical area worldwide other than the European Union, the United Kingdom and South Korea. Redistribution under Section 3.1 is confined to the same territory. This is the clause most likely to matter to a company that reads "open weights" and plans a deployment.

• The US$50 million line. Non-commercial use is free. Commercial use is also free while your annual revenue from related products or services is at or below US$50 million, provided you send a one-time use notice to Baidu before first commercial use and display the exact string "Powered by ERNIE" somewhere readily accessible — a UI, an about page, or, for APIs and backend services, your developer documentation. Above US$50 million, you need separate prior written authorisation before you start, and fees are set by a separate agreement.

• How revenue is counted. Section 1.7 aggregates every entity under common control — more than 50% of equity or voting rights — into one operator. It also counts the entire revenue of a related product even when that product uses other models alongside ERNIE. You cannot attribute your way under the threshold by pointing at the share of traffic that touches these weights.

The contrast with Baidu's own history is what makes this newsworthy rather than routine. ERNIE 4.5, released on June 30, 2025, shipped ten variants under Apache 2.0 — genuinely permissive, no territory, no revenue gate, no attribution string. ERNIE 5.0 in November 2025 went the other way entirely and shipped no weights at all. ERNIE-5.5-Flash-Base is a third position: weights you can download, under terms that make "open" a statement about availability rather than about permission.

If you are evaluating this model for anything commercial, read Sections 1.7, 1.11 and 4 before you read the benchmark table. The table tells you what it can do. Those three sections tell you whether you are allowed to find out.

What it takes to run

The weights are large enough that the hardware question answers itself, and small enough that the answer is not absurd. At FP8, 280B parameters is roughly 280 GB of weights before any KV cache — arithmetic, not a measurement — which fits a four-GPU node of 80 GB cards with room that is tighter than it sounds once you enable a 128K context. The BF16 repository doubles that to roughly 560 GB and wants eight cards.

Baidu ships two serving paths, and both carry friction worth knowing about before you start.

• SGLang — a prebuilt image, lmsysorg/sglang:dev-ernie-55-flash, launched with tensor parallelism 4 and attention data parallelism 4. It is a dev tag, not a pinned release.

• vLLM — vLLM 0.32.0 plus a plugin installed from Baidu's own vLLM-ERNIE repository. The launch flags include FP8 KV cache, expert parallelism, and --moe-backend humming.

Neither path is a one-line install, and the vLLM route in particular pins you to a specific vLLM version and a plugin that Baidu maintains. That is normal for a model with new attention and MoE structure, and it is the kind of detail that decides whether a two-week evaluation slips by a week. A Megatron-Bridge fine-tuning example is linked from the card for the same reason: this checkpoint is meant to be trained on, not just served.

Who should download it, and who should wait

Download it if you are set up to fine-tune. A 280B-A14B base checkpoint trained on 33.5T tokens, with a 1M-token inference window and commercial terms that are workable below the revenue line and outside the excluded territories, is a serious starting point for a domain adaptation or an RL run — and the activation ratio means your training and serving costs track 14B, not 280B. The efficiency work Baidu did on attention and KV cache is aimed squarely at the long-context, agentic workloads that are expensive to run on a dense model of comparable quality.

Wait if you want a model to call. There is no instruct version, no API, and no independent evaluation. Anyone who needs ERNIE 5.5 Flash as a product is waiting for the post-trained release Baidu has promised and not dated.

And check the map before either. If your users, your entity or your infrastructure sit in the EU, the UK or South Korea, the licence does not grant you the rights the download button implies.

There is a practical middle path for teams whose real question is not "can we run this" but "is this better than what we already pay for." The comparison you need is against the hosted models you would be replacing, and those you can call today. OrcaRouter puts 200-plus models behind one API at 0% markup — provider list price passed through, so a vendor price cut lands the same day — which makes a bake-off against your current production model a matter of changing a model string rather than signing a second contract. For a checkpoint you intend to fine-tune and self-host, the routing layer is how you keep the incumbent in the loop while you find out whether the fine-tune is actually better, with automatic failover if it is not. ERNIE-5.5-Flash-Base is not on OrcaRouter; it is a set of weights you run yourself, and nothing here changes that.

What to watch

The interesting variable is not whether the base checkpoint is good — on Baidu's own numbers it is competitive with the base models it was measured against, and it wins most of the rows. The variable is what Baidu does with the post-trained model, because that is where the licence terms start to bind on real products, and where an independent score will finally exist.

Two dates are worth holding onto. The first is the post-trained ERNIE 5.5 Flash, which Baidu says will follow and has not scheduled. The second is the first third-party reproduction of any row in that 30-row table — the moment the numbers stop being Baidu's and start being everyone's. Until then, the honest summary of ERNIE-5.5-Flash-Base is that Baidu shipped a large, carefully engineered pretraining checkpoint in public, described it accurately, and attached a licence that is considerably less open than the phrase "open weights" tends to imply.