Nothing here is sponsored — no affiliate links on this page. This breaking article is not the daily buyer-intent post. Facts below come from the sources listed at the end, checked 5 November 2026. We have not tested ERNIE 5.5 Flash Base hands-on, and nothing here is a hands-on finding.

Baidu's ERNIE 5.5 Flash Base cover image

Baidu has released a serious research artefact, and the honest way to describe it is also the least exciting-sounding one: a base model. On 5 November 2026, Baidu's ERNIE Team announced ERNIE-5.5-Flash-Base — a text-only mixture-of-experts model with 280 billion total parameters and 14 billion activated per token, with downloadable weights. The checkpoint ends at pre-training. There was no mid-training and no post-training. That is the point of the release, not a flaw in it — and it changes everything about who this model is for.

If you were hoping for a new chatbot to try this afternoon, stop here: this is not that. If you fine-tune models, run training research, or build on top of open weights, this is one of the most interesting releases of the month. The rest of this article is about why, and about the licence trap that could matter more than any benchmark.

How we put this together

We verified the story against the 5 November 2026 coverage — Unite.AI's report on the release, Artiverse's analysis, and the underlying model card and repositories they cite — keeping every benchmark claim labelled as Baidu's own and every licence condition tied to the licence text as reported. We have not run ERNIE 5.5 Flash Base ourselves, so nothing here is a hands-on finding. Where the announcement gives no number or date, we say so.

Verdict

ERNIE 5.5 Flash Base is a genuinely valuable release for researchers and fine-tuners: a large, modern MoE base model with downloadable weights, strong self-reported exam and coding scores, and a 1-million-token context ceiling. But it is a base model, not a product — it will not chat with you out of the box — and its custom licence is the real story for anyone building commercially: the EU, UK and South Korea are excluded entirely, and commercial use above $50 million in related annual revenue needs Baidu's written permission. Read the licence before you download the weights. The model is free to look at; it is not free of conditions.

What it is

ERNIE-5.5-Flash-Base is a text-only mixture-of-experts (MoE) base model. The two numbers that matter most are 280 billion and 14 billion: 280 billion total parameters, of which only about 14 billion are activated per token. That is the whole MoE bargain — a huge reservoir of specialised knowledge, with only a small slice doing work on any given token, which keeps inference costs far below what a dense 280-billion-parameter model would demand.

The architecture, as listed on the model card, is 43 layers with a hidden size of 4,096 and 64 attention heads. Two design choices are worth understanding even if you never touch the weights. First, 37 of the 43 layers use Heavily Compressed Attention: their cache holds one entry for each 128-token block, plus a short window of recent uncompressed tokens, while six Multi-head Latent Attention layers keep token-level access to the full context and use an indexer to pick the highest-scoring earlier positions. Baidu says this sparse selection keeps per-query core-attention cost constant as context grows — which is the engineering behind the long-context claim below. Second, every layer except the first uses an MoE block with 512 routed experts, of which 10 are activated, plus one shared expert. The experts operate in a 2,048-dimensional latent space rather than the model's full hidden width, which Baidu says roughly halves expert dispatch and combination traffic and expert FLOPs relative to full-width experts. The design also includes four parallel residual streams through manifold-constrained hyper-connections and a MultiMax output head with a learnable piecewise temperature. You do not need to memorise any of that. What it adds up to is a model built, from the ground up, to keep the cost of very long inputs under control.

Pre-training ran on 33.5 trillion tokens, using the Muon optimizer and FP8 mixed precision. The training context was gradually extended to 128K tokens, and fine-tuning and inference support up to 1 million tokens. The corpus is worth a look too, because it tells you what the model was built for: alongside multilingual web data it includes code and records of software development — commits, issues, pull requests, reviews, generated tests and CI-verified fixes — plus curated long documents, sequences with semantically similar distractors, and synthetic retrieval and question-answering data for long-context training. That is a corpus aimed squarely at repository-level reasoning, agentic software engineering and extended inputs. It is not a general chatbot diet. It is a working diet.

And the checkpoint itself is the key fact of this release: it ends at pre-training, before mid-training or post-training. Baidu is explicit that this gives developers a common starting point for domain adaptation, instruction tuning and reinforcement-learning research — and a fixed baseline for controlled comparisons of training methods. In other words, the value here is reproducibility. Everyone who fine-tunes this checkpoint starts from exactly the same place, and can compare methods against each other honestly. That is a genuinely useful thing for the research community, and it is also why this release will barely register with ordinary users: there is no chat interface, no instruction following, no alignment. A base model completes text. It does not answer your questions politely.

The headline numbers — labelled as Baidu's own

Baidu evaluated ERNIE against the publicly released DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base checkpoints using the same harness, the same prompts and few-shot examples, and the same LLM judge for free-form question answering. That is the right way to run a comparison, and it is worth saying so — but it is still Baidu's harness, Baidu's table, and Baidu's numbers. No independent lab has reproduced them yet.

Here is what the table shows, selected results:

  • MMLU, five-shot: ERNIE 93.3 — DeepSeek V4 Flash 89.4, Hy3 preview 87.5, MiMo V2.5 87.6
  • HumanEval+, one-shot Pass@1: ERNIE 82.3 — DeepSeek V4 Flash 69.5, Hy3 preview 78.0, MiMo V2.5 74.4
  • FullStackBench, one-shot Pass@1: ERNIE 54.6 — DeepSeek V4 Flash 43.3, Hy3 preview 56.7, MiMo V2.5 54.1
  • InfiniteBench, zero-shot average: ERNIE 46.5 — DeepSeek V4 Flash 48.0, Hy3 preview 33.0, MiMo V2.5 46.7

Read that as it is: a strong showing, not a clean sweep. ERNIE leads clearly on MMLU — a broad exam-style knowledge test — and on HumanEval+, a coding test, where its 82.3 against DeepSeek's 69.5 is a wide margin. But Tencent's Hy3 leads on FullStackBench, a broader coding benchmark, and DeepSeek leads on InfiniteBench, the long-context test — which is a little ironic, given how much of this model's design is aimed at long context. The honest summary is that ERNIE looks strong on exams and on single-function coding, competitive elsewhere, and not dominant across the board. Anyone reporting only the MMLU number is selling you a headline. The full table is the story.

One more caution about that MMLU figure: 93.3 is very high, and high MMLU scores on base models should always invite the same question — how much of it is the model, and how much is the evaluation setup? We cannot answer that from Baidu's table alone. Independent runs will tell. Until then, treat 93.3 as a claim to be checked, not a fact to be quoted.

What you can actually use today

The weights are downloadable in FP8 and BF16 safetensors formats. The model card documents production serving with SGLang and vLLM through the completions API — note: the completions API, matching the checkpoint's base-model status, not a chat API. The accompanying FlashMask repository documents attention kernels for Hopper and Blackwell GPUs, with Blackwell configurations supporting head dimensions up to 576 for query and key and 512 for value. Baidu reports speed gains of up to 175% in the forward pass and 74% in the backward pass over FlashAttention-4 MaskMod on SM100 hardware — Baidu's own figures, on Baidu's chosen comparison, so take the size of the gain with the usual pinch of salt while noting the direction is plausible for a purpose-built kernel.

There is also a practical on-ramp for fine-tuning: NVIDIA's Megatron Bridge recipe adds a supervised fine-tuning workflow covering data preparation, training and export back to Hugging Face format. Its full-model recipe uses 16 nodes with four GB300 GPUs per node, and NVIDIA validates the recipes on Blackwell GB200 and GB300 systems — the components support Hopper, but Hopper is not validated. Read that hardware line twice before planning a project around this model. A 280-billion-parameter model, even with only 14 billion active per token, is not a weekend laptop project for full fine-tuning. The weights are downloadable; running them well is a serious infrastructure commitment. Parameter-efficient fine-tuning may be the realistic route for most teams, and nothing in the sources for this article states a LoRA or adapter recipe, so we will not invent one.

Baidu says a post-trained ERNIE 5.5 Flash will follow this base version. No date has been given in the sources for this article. When it arrives, that is the version ordinary users and most developers will actually want — the one that follows instructions and chats. This base release is the foundation, not the house.

The licence trap — read this before you download

This is the section that matters most, and the one most coverage will bury. ERNIE 5.5 Flash Base is not released under Apache 2.0, MIT or any standard open licence. It is released under a custom ERNIE 5.5 Flash Base Model License Agreement v1.0, and its conditions are strict in ways that will exclude a large share of potential users outright.

  • Territory: the licence's territory excludes the European Union, the United Kingdom and South Korea. Use, modification and distribution outside the licensed territory are not authorized. If you are in the EU, the UK or South Korea, this model is, on the face of the licence, not for you.
  • Commercial use up to $50 million: non-commercial use is free of model licence charges. Commercial use at or below $50 million in annual revenue from related products or services requires a one-time notice and prominent "Powered by ERNIE" attribution.
  • Commercial use above $50 million: above that threshold, you need Baidu's separate prior written authorization. That is a negotiation, not a download.
  • Distillation and synthetic data: the agreement bars using the licensed materials, derivative models or their outputs to improve an unrelated AI or machine-learning system. That restriction matters for distillation and synthetic-data workflows, which the licence expressly addresses when defining derivative models and limits on output use. If your plan was to use ERNIE's outputs to train a smaller model, stop and read the licence first.

Put plainly: "downloadable weights" and "open" are not the same thing, and this release is the proof. Compare it with the recent open-weight releases that shipped under Apache 2.0 with no territorial carve-outs — this one is a different animal. For researchers inside the licensed territory doing non-commercial work, the conditions are light. For a European startup, a UK lab, a Korean company, or any business that might one day cross $50 million in related revenue, the conditions are a wall. That does not make the release worthless. It makes it conditional. Know which side of the conditions you are on before you invest weeks of fine-tuning work on top of a checkpoint you may not be allowed to ship.

Why it matters

Three things make ERNIE 5.5 Flash Base worth attention even before independent verification. First, a fixed, downloadable, pre-training-only checkpoint at this scale is genuinely useful infrastructure for research: it lets the community test training methods, alignment techniques and long-context approaches against a common starting point, without anyone having to repeat a 33.5-trillion-token pre-training run. Second, the architecture is a serious attempt at the problem every frontier lab is wrestling with — how to make very long context affordable — and the combination of compressed attention, latent-space experts and a purpose-built kernel is a design worth studying whether or not you ever run the model. Third, the release sharpens a question the open-weight world has been avoiding: what "open" means when the licence carries territorial exclusions and revenue gates. Each release like this forces the community to be more precise about the difference between downloadable weights and genuinely open weights. That precision is overdue.

Who it matters for

  • Researchers in training methods and alignment. A fixed base checkpoint at 280B scale, with a documented training recipe lineage and a post-trained version coming, is exactly the baseline you want for controlled comparisons. Check the licence territory first.
  • Teams fine-tuning for code and agentic software engineering. The corpus is built for repository-level work, the HumanEval+ number is strong, and SGLang/vLLM serving plus a Megatron Bridge SFT recipe give you a real starting point — if you have the hardware and the licence allows your use.
  • Developers who just want a model to chat with. Wait for the post-trained ERNIE 5.5 Flash. This base checkpoint will not do that job.
  • Anyone in the EU, UK or South Korea, or building a commercial product that may cross $50M in related revenue. Read the licence before anything else. The territorial exclusion and the written-authorization threshold may rule this model out for you entirely.
  • Anyone tracking the open-weight race. ERNIE 5.5 Flash Base is another data point that the race is no longer just about parameter count: training data composition, attention design for long context, and — increasingly — licence terms matter as much as size.

The limits — stated plainly

  • This is a base model, not a finished assistant. No instruction tuning, no alignment, no chat behaviour. It completes text; it does not follow instructions.
  • All benchmark claims are Baidu's own, with no independent verification yet. MMLU 93.3 and HumanEval+ 82.3 are self-reported, and the comparison is task-dependent — Hy3 leads on FullStackBench, DeepSeek leads on InfiniteBench.
  • The licence is custom and restrictive. EU, UK and South Korea excluded; commercial use above $50 million in annual related revenue needs Baidu's prior written authorization; outputs may not be used to improve an unrelated AI system. This is not Apache 2.0.
  • Full fine-tuning needs serious hardware. NVIDIA's full-model SFT recipe uses 16 nodes with four GB300 GPUs each, validated on GB200/GB300; Hopper is supported but not validated.
  • The post-trained ERNIE 5.5 Flash has no date. Baidu says it will follow; no date has been given in the sources for this article.
  • The 1-million-token context applies to fine-tuning and inference, not to pre-training, which extended to 128K tokens. Treat hosted long-context behaviour as claimed until providers confirm it.
  • No hands-on testing was performed for this article, and no benchmarks, prices or dates have been invented.

The bottom line is a measured one. ERNIE 5.5 Flash Base is a serious, useful release for the people it is actually for — researchers and fine-tuners working inside its licence territory, with the hardware to run a 280-billion-parameter model and the patience to build on a raw checkpoint. Its exam and coding numbers are loud, its long-context engineering is thoughtful, and its fixed-baseline approach is genuinely good for science. But the licence is the story: a model that excludes the EU, the UK and South Korea, gates large-scale commercial use behind Baidu's written permission, and restricts distillation is not "open" in the way most developers mean the word. Download the weights if the terms fit you. If they don't, admire the engineering from a distance — and wait for the post-trained version, or for a release whose licence matches its ambitions.

Sources: Unite.AI, "Baidu Releases ERNIE 5.5 Flash Base for Further Training", 5 November 2026; Artiverse, "Baidu's ERNIE 5.5 Base Targets Serious Model Research"; Baidu ERNIE Team announcement and model card; FlashMask repository; NVIDIA Megatron Bridge recipe; ERNIE 5.5 Flash Base Model License Agreement v1.0. Benchmark figures are Baidu's own, as reported in the coverage. No hands-on testing was performed for this article. This breaking article is not the daily buyer-intent post, and it contains no affiliate links.

Originally published at ToolsDig