ERNIE 5.5 Flash Base Explained: Baidu’s 280B Open Model, Benchmarks & Licence Limits
On 5 November 2026, Baidu released ERNIE 5.5 Flash Base, the raw pre-training checkpoint of its newest large language model — and it is one of the most interesting open-weight releases of the year. The model carries 280 billion parameters in total but activates only 14 billion of them for each token it processes, it was trained on 33.5 trillion tokens, and it can handle a context window of up to one million tokens once fine-tuned. Baidu published the full weights in FP8 and BF16 on Hugging Face, free to download, alongside a 30-benchmark comparison in which it beats its closest open-weight rivals on most tests. There is a catch, and it is a big one: this is a base model, not a chatbot. It cannot follow instructions or hold a conversation out of the box, and its licence excludes the European Union, the United Kingdom and South Korea entirely. This guide explains what ERNIE 5.5 Flash Base actually is, how its hybrid-attention architecture works, where it wins and loses on benchmarks, what hardware you need to run it, and who should — and should not — build on it.

Table of Contents
- ERNIE 5.5 Flash Base at a Glance
- What Is ERNIE 5.5 Flash Base?
- Inside the Architecture: 280B Parameters, 14B Active
- How ERNIE 5.5 Flash Base Was Trained
- ERNIE 5.5 Flash Base Benchmarks: Where It Wins and Loses
- Who Is ERNIE 5.5 Flash Base Actually For?
- Running ERNIE 5.5 Flash Base: Hardware and Setup
- The Licence: Read This Before You Download
- ERNIE 5.5 Flash Base Pricing: What Does It Cost?
- ERNIE 5.5 Flash Base vs DeepSeek, Hy3 and MiMo
- How to Download and Run ERNIE 5.5 Flash Base
- Limitations and Honest Caveats
- What Comes Next: The Post-Trained ERNIE 5.5 Flash
- ERNIE 5.5 Flash Base FAQs
- The Bottom Line
ERNIE 5.5 Flash Base at a Glance
If you only have thirty seconds, here is the whole story. ERNIE 5.5 Flash Base is Baidu’s 280-billion-parameter mixture-of-experts language model, released as open weights on 5 November 2026. Only 14 billion parameters switch on for any given token, which is what makes a model this large affordable to run. It is a base checkpoint — the weights exactly as they stood when pre-training finished, before any instruction tuning or reinforcement learning — so it is raw material for developers and researchers, not a finished assistant. The headline numbers are strong: Baidu’s own evaluation shows it leading rival base models on knowledge, coding and long-context tests, including a 93.3 on MMLU and an 82.3 on HumanEval+. The weights are free to download, there is no hosted API price yet, and the licence carries real restrictions on where you can use it and what you can do with its outputs.
| Specification | ERNIE 5.5 Flash Base |
|---|---|
| Released | 5 November 2026 |
| Developer | Baidu ERNIE Team |
| Total parameters | 280 billion |
| Active parameters per token | 14 billion |
| Architecture | Mixture-of-experts (MoE), hybrid attention |
| Layers | 43 (1 dense + 42 MoE) |
| Pre-training data | 33.5 trillion tokens |
| Context length | 128K in training; 1M for fine-tuning and inference |
| Modality | Text only |
| Weights | FP8 and BF16 safetensors on Hugging Face and ModelScope |
| Serving | SGLang and vLLM (completions API) |
| Licence | ERNIE 5.5 Flash Base Model License Agreement v1.0 |
| Price | Free to download; no hosted API yet |
What Is ERNIE 5.5 Flash Base?
To understand ERNIE 5.5 Flash Base, it helps to know what a base model is — because that single fact shapes everything else in this article. When a lab trains a large language model, the first and longest phase is pre-training: the model reads trillions of tokens of text and learns to predict what comes next. The result is a base model. It is very good at continuing text, completing code and recognising patterns, but it has never been taught to answer questions, follow instructions or refuse a bad request. That teaching happens later, in mid-training and post-training, through instruction tuning and reinforcement learning.
Baidu stopped the process at the end of pre-training and published exactly that checkpoint. The company’s own model card says it plainly: ERNIE 5.5 Flash Base is “the checkpoint at the end of pre-training, before any mid- or post-training”. That is a deliberate choice, not an unfinished release. Different teams need different things from a foundation model — a legal-tech company wants a model tuned on contracts, a coding startup wants one tuned on repositories, a research lab wants to test a new training method against a fixed starting point. By releasing the raw checkpoint, Baidu gives all of them the same starting line instead of forcing them to undo someone else’s post-training first.
The trade-off is that ordinary users get nothing usable directly. You cannot open a chat window and talk to ERNIE 5.5 Flash Base. It is served through a completions endpoint, not a chat endpoint: you send it a piece of text and it continues that text. Ask it a question in the style of a chat prompt and it may answer, or it may simply invent more questions in the same style. That behaviour is normal for a base model, and it is why Baidu says a post-trained ERNIE 5.5 Flash — the finished, chat-ready version — “will follow”. No date has been given.
This release also matters because of who Baidu is in the open-weight race. The company behind China’s most-used search engine has been steadily open-sourcing its ERNIE line, and ERNIE 5.5 Flash Base lands in a month crowded with open-weight releases from Chinese labs. The company behind China’s most-used search engine has been steadily open-sourcing its ERNIE line, and this release lands in a month crowded with open-weight releases from Chinese labs. Tencent’s Hy4, which we covered earlier, is one of the three models Baidu benchmarks against. DeepSeek’s V4 Flash is another. The pattern is clear: the open-weight frontier is no longer a single lab’s story, and each release now competes on efficiency — how much capability you get per active parameter — rather than on raw size alone.
What ERNIE 5.5 Flash Base Is Not
Three common misunderstandings are worth clearing up before you go further. First, ERNIE 5.5 Flash Base is not a chatbot, a coding assistant or an app of any kind. There is no website where you can try it, no mobile app and no hosted API with a price per token. Second, it is not a multimodal model. It handles text only — no images, no audio, no video. Third, it is not a small model you can run on a laptop. The FP8 weights alone take about 288 GB of disk space, and serving the model needs multiple server GPUs. Anyone telling you to “just download it and chat” has not read the model card.
None of that makes the release unimportant. Base-model releases are the raw material the whole open ecosystem builds on, and a 280-billion-parameter checkpoint with strong benchmark scores and a million-token context is serious raw material. The rest of this article is about judging it on those terms.

Inside the Architecture: 280B Parameters, 14B Active
The core idea behind ERNIE 5.5 Flash Base is efficiency, and the architecture is where Baidu spent its effort. The model is a sparse mixture-of-experts design: instead of one giant network that fires on every token, it contains hundreds of smaller “expert” networks, and a routing system picks a handful of them for each token. The total parameter count is 280 billion, but only 14 billion are active per token — roughly 5% of the model does the work at any moment. That is the whole point of the “Flash” name: big capability, small active cost.
The details, taken from the official model card, show how carefully Baidu tuned each component:
- 43 layers in total. The first layer is a dense feed-forward layer; the other 42 are mixture-of-experts layers. Each MoE layer holds 512 routed experts, of which the top 10 are activated for a token, plus one shared expert that is always on.
- Half-width latent experts. The experts operate in a 2,048-dimensional latent space rather than the model’s full 4,096-dimensional hidden width. Baidu says that roughly halves the traffic of sending data to experts and combining their outputs, and halves expert computation, compared with full-width experts.
- Multiple routers, not one. A single router’s scores correlate only weakly with which expert is actually useful, especially in the middle and deep layers. ERNIE 5.5 Flash Base uses several independently initialised sigmoid routers with different preferences, so expert selection no longer depends on one router’s mistakes — without changing the number of activated experts or the computation cost.
- Auxiliary-loss-free load balancing. The experts stay evenly loaded without the extra loss term that older MoE models needed, which keeps training stable. Each routed expert’s routing weight is also multiplied by a learnable scalar of its own.
Hybrid Attention: The Clever Part
The most distinctive piece of ERNIE 5.5 Flash Base is its attention system, because long context is where large models usually get expensive. Attention is the mechanism that lets a model look back at earlier tokens, and its cost normally grows with the length of the context. Baidu’s answer is a hybrid: 37 of the 43 layers use Heavily Compressed Attention (HCA), and 6 layers use Multi-head Latent Attention (MLA) with sparse attention.
The HCA layers are the workhorse. Each one stores a single cache entry for every 128-token block, plus a short window of recent uncompressed tokens. That compresses the key-value cache — the memory a model needs to remember what it has read — by roughly 128 times for most layers. The MLA layers, one in every seven, keep full token-level access to the whole context and use an indexer to pick the highest-scoring earlier positions. Baidu says this sparse selection keeps the per-query core attention cost constant as the context grows, which is what makes the million-token context practical rather than theoretical.
Two more refinements sit inside every attention layer. Low-Rank Gated Attention computes the attention output gate from a low-rank query latent instead of the hidden state, cutting gate-projection parameters by 75% while improving performance. Virtual Head Attention adds a lightweight low-rank mixing step across attention heads, fused into the attention operator at inference time. Neither is glamorous on its own; together they are the kind of compounding efficiency that defines this model.
Hyper-Connections, MultiMax and ECHO
Three more design choices round out ERNIE 5.5 Flash Base. First, instead of the standard residual stream that carries information between layers, the model uses manifold-constrained hyper-connections with an expansion rate of 4 — four parallel residual streams — which Baidu says strengthens signal propagation across all 43 layers. Second, the output head is a MultiMax head with a learnable piecewise temperature, which sharpens the model’s vocabulary distribution at almost no extra training cost. Third, the model includes ECHO, a mechanism that reuses the backbone’s final layer to predict future tokens instead of adding a separate multi-token-prediction module. That has a practical bonus: the same layer doubles as a built-in draft model for speculative decoding, which can speed up text generation at inference time.
The tokenizer is a roughly 200,000-entry byte-pair encoding vocabulary that tokenises numbers right-to-left in groups of up to three digits — a small detail that helps the model handle arithmetic and long numbers more reliably than left-to-right digit splitting.
How ERNIE 5.5 Flash Base Was Trained
ERNIE 5.5 Flash Base was pre-trained from scratch on 33.5 trillion tokens, using the Muon optimiser and FP8 mixed precision. For scale, 33.5 trillion tokens is roughly in the same league as the largest public pre-training runs disclosed this year, and the choice of Muon — a newer optimiser that has been gaining ground over AdamW in large-scale training — signals that Baidu’s training team is working at the current frontier of training practice, not reusing a five-year-old recipe.
The training data was chosen with the model’s intended uses in mind. Alongside large-scale multilingual web data, filtered for quality by model-driven grading, the corpus includes a heavy dose of code and software-development records: repository code, commits, issues, pull requests, code reviews, requirement-to-PR trajectories, generated tests and CI-verified fixes. That is not accidental. Baidu explicitly aligned the data with repository-level reasoning and agentic software engineering — the model is meant to be fine-tuned into a coding agent, so it was fed the raw material of real software work. Long-context training came from curated long documents, sequences built with semantically similar distractors, and synthetic retrieval, citation-grounded question answering and multi-hop QA data.
The context length was extended gradually during training to 128,000 tokens. Fine-tuning and inference support up to 1,048,576 tokens — a full million. That matters because the model was designed for the long-document, whole-repository workloads that developers actually have, not just for passing long-context benchmarks.
Baidu also co-designed the training system with the architecture. The distributed Muon setup fuses dense and MoE parameter synchronisation into a single collective operation instead of a cluster-wide ring, cutting inter-node latency and jitter across thousands of GPUs without changing the optimiser’s maths. Optimiser state is spread nearly evenly across nodes, removing the out-of-memory hot spots and slow checkpointing that plague very large runs. And the team released FlashMask-4, a new attention kernel for Nvidia Hopper and Blackwell GPUs, which Baidu reports is up to 175% faster in the forward pass and 74% faster in the backward pass than FlashAttention-4 MaskMod on SM100 hardware. FlashMask is published as its own repository, so the kernel work benefits the wider community even if you never touch the model.

ERNIE 5.5 Flash Base Benchmarks: Where It Wins and Loses
Benchmarks are where ERNIE 5.5 Flash Base makes its case — and where honest reading matters most. Baidu evaluated the model against three publicly released base checkpoints of similar size, using the same harness, the same prompts, the same few-shot settings and the same LLM judge for free-form answers: DeepSeek-V4-Flash-Base (284B total, 13B active), Hy3-preview-Base (295B total, 21B active) and MiMo-V2.5-Base (310B total, 15B active). All four models are base checkpoints, so the comparison is apples to apples. Every number below is Baidu’s own, reported on the official model card; no independent lab has re-run the evaluation yet, and that caveat applies to everything in this section.
The headline results are genuinely strong. On general knowledge, ERNIE 5.5 Flash Base scores 93.3 on MMLU (five-shot), ahead of DeepSeek’s 89.4, Hy3’s 87.5 and MiMo’s 87.6, and 78.4 on the much harder MMLU-Pro, where the nearest rival manages 69.2. On coding, the lead is wider still: HumanEval+ comes in at 82.3 against DeepSeek’s 69.5, MBPP+ at 81.5 against 67.2, and LiveCodeBench at 40.1 against 31.9. Long-context results are strong too, with RepoQA at 91.7 — well ahead of DeepSeek’s 81.0 and MiMo’s 56.5 — and LongBench-v2 at 48.7, the best of the four.
| Benchmark | ERNIE 5.5 Flash Base | DeepSeek V4 Flash | Hy3 preview | MiMo V2.5 |
|---|---|---|---|---|
| MMLU (5-shot) | 93.3 | 89.4 | 87.5 | 87.6 |
| MMLU-Pro (5-shot) | 78.4 | 69.2 | 66.9 | 65.7 |
| GPQA (5-shot) | 60.2 | 55.9 | 50.2 | 51.4 |
| HumanEval+ (1-shot) | 82.3 | 69.5 | 78.0 | 74.4 |
| LiveCodeBench (1-shot) | 40.1 | 31.9 | 33.5 | 32.4 |
| FullStackBench (1-shot) | 54.6 | 43.3 | 56.7 | 54.1 |
| RepoQA (0-shot) | 91.7 | 81.0 | 91.3 | 56.5 |
| InfiniteBench (0-shot avg) | 46.5 | 48.0 | 33.0 | 46.7 |
| ChineseSimpleQA (10-shot) | 73.0 | 75.3 | 73.2 | 68.8 |
| BBH chain-of-thought (3-shot) | 89.2 | 87.5 | 90.0 | 89.8 |
But it is not a clean sweep, and the exceptions are instructive. Hy3-preview-Base wins FullStackBench (56.7 vs 54.6) and BBH with chain-of-thought prompting (90.0 vs 89.2). DeepSeek-V4-Flash-Base wins InfiniteBench (48.0 vs 46.5), ChineseSimpleQA (75.3 vs 73.0), INCLUDE-lite and HELMET-Rerank. Across the full 30-benchmark table, ERNIE 5.5 Flash Base leads on most tests, but the losses cluster in long-context reranking and a few Chinese-language factuality tests — worth knowing if those are your workloads.
Three Benchmark Caveats You Should Not Skip
First, the comparison targets are not each vendor’s latest models. Baidu compared against DeepSeek-V4-Flash-Base, but DeepSeek has since shipped the newer V4.1-Flash-1017; against Hy3-preview-Base, while Tencent’s Hy4 is already public; and against MiMo-V2.5-Base, while MiMo-V2.6-Flash exists. The rivals in the table are the latest publicly released base checkpoints Baidu could test under one harness, but the comparison is against last month’s field, not today’s.
Second, these are base-model benchmarks, not assistant benchmarks. A base model that scores 93.3 on MMLU will not necessarily produce the best chatbot after fine-tuning — post-training quality, safety tuning and tool use all change the picture. Judge ERNIE 5.5 Flash Base as raw material, and wait for the post-trained release before comparing it with finished assistants like GPT-5 or Claude.
Third, the evaluation was run by Baidu itself, on its own harness, with its own choice of prompts and an LLM judge for free-form answers. That does not make the numbers wrong — the methodology is disclosed and the rivals were tested identically — but it does mean independent reproduction is still pending. Treat the scores as a strong signal, not a settled fact.
Who Is ERNIE 5.5 Flash Base Actually For?
The honest answer is that ERNIE 5.5 Flash Base is for a narrow audience, and Baidu designed it that way. If you are in one of these groups, the release is genuinely useful. If you are not, you should wait for the post-trained version.
- AI research labs. A fixed, public 280B checkpoint is a shared baseline. Teams can test new fine-tuning methods, reinforcement-learning recipes or evaluation harnesses against the same starting point without repeating a 33.5-trillion-token pre-training run. That is the use Baidu names first in its announcement, and it is the most credible one.
- Companies building domain-specific models. A legal-tech firm, a medical-documentation startup or a financial-analysis team can take ERNIE 5.5 Flash Base and fine-tune it on their own data, ending up with a model that speaks their domain fluently. The million-token context is a real advantage here: whole contracts, whole patient files or whole codebases fit in one pass.
- Teams building coding agents. The training data leans heavily on repository-level software work — commits, pull requests, reviews, CI-verified fixes — and the code benchmarks are the model’s strongest results. If you are building an agent that reads whole repositories, this checkpoint is a serious starting point.
- Infrastructure and serving teams. SGLang and vLLM support is documented from day one, the FlashMask-4 kernel is public, and Nvidia ships a Megatron Bridge fine-tuning recipe. Teams that run their own GPU clusters can evaluate the model on their own hardware immediately.
And who should skip it? Everyone looking for a chatbot, a writing assistant or a plug-and-play API. There is no chat interface, no hosted endpoint and no price per token. Hobbyists with a single GPU should skip it too — the FP8 weights need about 288 GB, and no GGUF or MLX build exists yet for consumer hardware. If you want to use an ERNIE model rather than train one, the post-trained ERNIE 5.5 Flash, when it arrives, is the release to wait for.

Running ERNIE 5.5 Flash Base: Hardware and Setup
Running ERNIE 5.5 Flash Base is a data-centre job, and Baidu is upfront about that. The FP8 checkpoint is about 288 GB — roughly half the storage of the largest Mac Studio, as one early analysis noted — but storage is not the binding constraint. The model is served with SGLang or vLLM on server GPUs, and the official recipes assume multi-GPU setups: the SGLang example runs with tensor parallelism across 4 GPUs, and Nvidia’s full-model fine-tuning recipe uses 16 nodes with four GB300 GPUs each. That is 64 Blackwell GPUs for full fine-tuning, which puts full-model training firmly in the realm of companies and well-funded labs.
For inference, the picture is more accessible. The model card documents production serving through SGLang (with a prebuilt Docker image) and vLLM (version 0.32.0 with Baidu’s vLLM-ERNIE plugin), both exposing a completions endpoint on port 30000. A simple curl request to /v1/completions with a text prompt is enough to verify the server is running. Both servers accept contexts up to 1,048,576 tokens for long-context inference, though the SGLang example configures 131,072 by default.
What you cannot do yet is run ERNIE 5.5 Flash Base on consumer tooling. As of the release date, neither llama.cpp nor MLX supports the model’s new ernie5_5_moe architecture, and no GGUF or MLX build exists on Hugging Face. That will probably change — community ports of popular architectures usually arrive within weeks — but today the model belongs on server GPUs with SGLang or vLLM, not on a laptop. If your plan was to run it locally on a Mac, wait for community support or for a quantised community build.
One more practical note: the model is text-only, and the tokenizer is a 201,216-entry vocabulary. If you are budgeting context, remember that the 128K training context and the 1M inference context are different things — the model was trained with contexts up to 128K and the longer window is enabled for fine-tuning and inference, which is standard practice but worth knowing when you test very long inputs.
The Licence: Read This Before You Download
The licence is the part of the ERNIE 5.5 Flash Base release that most coverage glosses over, and it is the part most likely to affect your plans. The model is released under Baidu’s own ERNIE 5.5 Flash Base Model License Agreement v1.0 — a custom agreement, not a standard open-source licence like Apache 2.0 or MIT. Four terms matter.
- Territory exclusion. The licence covers the whole world except the European Union, the United Kingdom and South Korea. Use, modification and distribution in those territories are not authorised under the agreement. That is unusual — most open-weight licences have no geographic restriction at all — and it means a developer in Berlin, London or Seoul cannot legally use the model under this licence, even though the weights are downloadable from Hugging Face without any access form or geo-check.
- Commercial use up to $50 million. Non-commercial use is free of licence charges. Commercial use is also free of charges if your annual revenue from related products or services is at or below $50 million — but it requires a one-time notice to Baidu and prominent “Powered by ERNIE” attribution. Above $50 million in related annual revenue, you need Baidu’s separate prior written authorisation.
- No capability transfer to unrelated models. 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: you cannot use ERNIE 5.5 Flash Base to train a competing model, or to generate training data for one, outside the terms the licence sets for derivative models.
- Attribution is not optional. The “Powered by ERNIE” notice is a condition of the free commercial tier, not a courtesy. If you ship a product built on the model, the attribution has to be visible.
For a US-based developer or small company, the licence is workable: free below $50 million in related revenue, with a notice and attribution. For a European team, it is a hard stop unless Baidu changes the terms. And for anyone planning to distil the model into a smaller one or use its outputs to train something else, the restriction on improving unrelated systems needs a careful read of the full agreement before you start. The licence text is published alongside the model on Hugging Face — read it before you build, not after.

ERNIE 5.5 Flash Base Pricing: What Does It Cost?
There is no price list for ERNIE 5.5 Flash Base, and that is the honest answer. The weights are free to download from Hugging Face and ModelScope, in FP8 and BF16, with no licence fee for non-commercial use and no fee for commercial use at or below $50 million in related annual revenue (subject to the notice and attribution conditions above). There is no hosted API, so there is no per-token price to quote — no $X per million input tokens, no subscription tier, nothing to compare with the API pricing of DeepSeek or OpenAI.
The real cost is infrastructure. Storing 288 GB of FP8 weights is cheap; serving them is not. You need multiple server GPUs with enough memory to hold the model and its key-value cache, and the million-token context means the cache can grow very large on long inputs. Full-model fine-tuning, per Nvidia’s recipe, assumes 64 GB300 GPUs. For most teams, the practical route is to rent GPU time from a cloud provider and treat the model as a fixed cost of a project, not as a per-query expense.
That is also why the efficiency story matters commercially. Only 14 billion parameters activate per token, so inference compute per token is far lower than a dense 280-billion-parameter model would need. Baidu’s design goal was to make a very large model cheap to run once it is loaded — and on that measure, the architecture delivers. Whether that translates into lower bills for your workload depends on your serving setup, your batch sizes and your context lengths, but the active-parameter count is the right number to watch.
ERNIE 5.5 Flash Base vs DeepSeek, Hy3 and MiMo
Placed beside its three benchmark rivals, ERNIE 5.5 Flash Base looks like the strongest all-rounder of the group — with the important qualifier that all four are base checkpoints and the comparison is Baidu’s own. DeepSeek-V4-Flash-Base is the closest rival on long-context reranking and Chinese factuality; Hy3-preview-Base edges it on full-stack coding and chain-of-thought reasoning; MiMo-V2.5-Base trails on most tests despite having the largest total parameter count at 310 billion. The lesson is one the whole open-weight field keeps relearning: total parameters are a poor predictor of quality. ERNIE 5.5 Flash Base has the smallest total parameter count of the four and the best overall scores, because its architecture spends its parameters more carefully.
For context, the site has covered several of these rivals before. Tencent’s Hy4 is the post-trained successor to the Hy3 preview model in Baidu’s table, and it is worth reading alongside this article if you want the full Chinese open-weight picture. Kimi K3.1 Pro is another recent open-weight release with downloadable weights, and its coverage explains the practical side of running large open models. And if you want the efficiency story at the opposite end of the scale, TwIL-8B shows what an 8-billion-parameter model can do — a useful contrast to a 280-billion-parameter one. The open-weight cluster is moving fast, and these models are best understood together rather than alone.
| Model | Total / active params | Context | MMLU | HumanEval+ | Licence |
|---|---|---|---|---|---|
| ERNIE 5.5 Flash Base | 280B / 14B | 1M | 93.3 | 82.3 | Custom (EU/UK/KR excluded) |
| DeepSeek V4 Flash Base | 284B / 13B | — | 89.4 | 69.5 | Open-weight |
| Hy3 preview Base | 295B / 21B | — | 87.5 | 78.0 | Open-weight |
| MiMo V2.5 Base | 310B / 15B | — | 87.6 | 74.4 | Open-weight |
The comparison also underlines a strategic point. Baidu is not trying to win on size — it is trying to win on efficiency and on giving developers a clean starting point. The hybrid-attention design, the half-width latent experts, the multi-router system and the FlashMask kernel are all aimed at the same goal: a model that costs less to train, less to serve and less to fine-tune than its rivals. If that efficiency holds up under independent testing, ERNIE 5.5 Flash Base will be remembered as a design win, not just a benchmark win.
How to Download and Run ERNIE 5.5 Flash Base
If you have the hardware and you are in a territory the licence covers, getting started is straightforward. The official Hugging Face model card for ERNIE 5.5 Flash Base is the authoritative source, and the steps below follow its deployment section.
- Check the licence first. Confirm that your location is inside the licence territory — the EU, UK and South Korea are excluded — and that your planned use fits the commercial terms if you are building a product.
- Download the weights. The FP8 checkpoint (about 288 GB) is the practical choice for serving; the BF16 checkpoint is larger and suited to fine-tuning. Both are available on Hugging Face and ModelScope.
- Pick a serving engine. SGLang, with Baidu’s prebuilt Docker image, or vLLM 0.32.0 with the vLLM-ERNIE plugin. Both are documented on the model card with copy-paste commands.
- Launch the server. The SGLang example uses tensor parallelism across 4 GPUs and a 131,072-token context by default; both servers accept contexts up to 1,048,576 tokens.
- Send a completion request. As a base model, ERNIE 5.5 Flash Base is served through the completions endpoint, not chat. Send a text prompt and a token limit, and the model continues the text.
- Fine-tune before you ship. A raw base model is not a product. Plan an instruction-tuning or domain-adaptation step — Nvidia’s Megatron Bridge recipe covers supervised fine-tuning end to end — before putting the model in front of users.
If you are evaluating rather than deploying, the fastest path is the SGLang Docker image, which bundles the model support and gets a server running without building anything from source. Budget a few hours for the 288 GB download on a fast connection, and remember that you will need a machine with multiple server GPUs — this is not a job for a single consumer card.
Limitations and Honest Caveats
No model release is without limits, and ERNIE 5.5 Flash Base has several that deserve plain statement.
- It is not usable out of the box. A base model needs instruction tuning before it behaves like an assistant. Budget that work — data, compute and evaluation — into any plan that uses it.
- All benchmarks are vendor-reported. Baidu ran the evaluation, chose the harness and published the table. Independent reproduction is still pending, and the rivals in the table are not each vendor’s newest models.
- The licence excludes three major regions. The EU, UK and South Korea are outside the licence territory. That is a hard constraint for any team based there, whatever the benchmarks say.
- No consumer-GPU support yet. llama.cpp and MLX do not support the new architecture, and no GGUF or MLX build exists. Local running on a laptop or Mac is not possible today.
- Full fine-tuning needs serious hardware. Nvidia’s full-model recipe assumes 64 GB300 GPUs. Smaller teams will need parameter-efficient methods or hosted fine-tuning, and should test whether those methods preserve the model’s benchmark strengths.
- Distillation and synthetic-data use are restricted. The licence bars using the model’s materials or outputs to improve an unrelated AI system, which limits common workflows like distilling a large model into a small one.
- Long-context claims need your own testing. The 1M-token window is enabled for fine-tuning and inference, but the model was trained with contexts up to 128K. Very long inputs should be validated on your own data, not assumed to work at full quality.
None of these is a reason to dismiss the release. They are reasons to use it for what it is — a strong, efficient research and development checkpoint — and to wait for the post-trained version if you want a finished assistant.
What Comes Next: The Post-Trained ERNIE 5.5 Flash
Baidu has said a post-trained ERNIE 5.5 Flash will follow the base release, with no date given. That version is the one most people are actually waiting for: instruction-tuned, able to hold a conversation, and presumably available through Baidu’s own services with a price. When it lands, the questions to ask are different from the ones in this article — how good is the chat quality, what does the API cost, what are the safety and refusal behaviours, and does the licence still exclude the EU, UK and South Korea?
Two other things are worth watching. First, whether independent labs reproduce Baidu’s benchmark numbers — the open-weight community is quick to test new releases, and a confirming result would strengthen the efficiency story considerably. Second, whether llama.cpp, MLX or another community project adds support for the ernie5_5_moe architecture, which would open the model to far more developers. Both usually happen within weeks of a major release, and either would change who can practically use ERNIE 5.5 Flash Base.
The wider context is that the open-weight race is now a rolling monthly event. Kimi, Tencent, DeepSeek, MiMo and now Baidu have all shipped major open models in the past few weeks, each competing on a different axis — size, efficiency, context length, coding skill. FLUX 3 Image Klein showed the same pattern in image generation: open weights, efficient design, self-hosting as the pitch. The trend is unmistakable, and ERNIE 5.5 Flash Base is Baidu’s entry in it.
ERNIE 5.5 Flash Base FAQs
Is ERNIE 5.5 Flash Base free to use?
The weights are free to download, and non-commercial use is free of licence charges. Commercial use is free of charges up to $50 million in annual revenue from related products or services, provided you send Baidu a one-time notice and display “Powered by ERNIE” attribution. Above that threshold you need Baidu’s prior written authorisation. There is no hosted API and no per-token price yet.
Can I chat with ERNIE 5.5 Flash Base?
No — not directly. It is a base model served through a completions endpoint, so it continues text rather than answering questions. You need to fine-tune it with instruction data first, or wait for the post-trained ERNIE 5.5 Flash that Baidu says will follow.
How big is ERNIE 5.5 Flash Base?
The model has 280 billion total parameters with 14 billion active per token. The FP8 weights take about 288 GB of disk space. It was pre-trained on 33.5 trillion tokens and supports a context of up to 1 million tokens for fine-tuning and inference.
Can I run ERNIE 5.5 Flash Base on my laptop?
Not today. It needs multiple server GPUs with SGLang or vLLM, and neither llama.cpp nor MLX supports its architecture yet. No GGUF or MLX build exists. Community ports may arrive, but at launch this is a data-centre model.
Can I use ERNIE 5.5 Flash Base in Europe or the UK?
Not under the current licence. The ERNIE 5.5 Flash Base Model License Agreement v1.0 excludes the European Union, the United Kingdom and South Korea from its territory, so use, modification and distribution there are not authorised — even though the weights download without a geo-check.
How does ERNIE 5.5 Flash Base compare with DeepSeek?
On Baidu’s own evaluation, ERNIE 5.5 Flash Base beats DeepSeek-V4-Flash-Base on most of 30 benchmarks, including MMLU (93.3 vs 89.4) and HumanEval+ (82.3 vs 69.5). DeepSeek wins a few tests, including InfiniteBench and ChineseSimpleQA. Note the comparison is against the older V4 Flash base checkpoint, not the newer V4.1-Flash-1017, and all figures are vendor-reported.
The Bottom Line
ERNIE 5.5 Flash Base is a serious release that deserves to be judged on its own terms. As a piece of engineering, it is impressive: a 280-billion-parameter model that activates only 14 billion per token, a hybrid attention system that makes a million-token context practical, and benchmark results that lead its closest open-weight rivals on knowledge, coding and long-context tests. As a piece of open infrastructure, it is genuinely useful: researchers get a fixed baseline, domain teams get a strong starting point for fine-tuning, and the FlashMask kernel and Megatron Bridge recipe benefit the wider community. And as a product, it is not one — it is raw material, and Baidu is honest about that.
The two things that will decide how far it goes are outside Baidu’s benchmark table. The first is the licence: excluding the EU, UK and South Korea from the territory is a real limitation that no benchmark score offsets, and the restriction on using outputs to improve unrelated models will give pause to teams whose workflow is distillation. The second is what happens next: whether independent testing confirms the numbers, whether community tooling brings the model to more hardware, and when the post-trained ERNIE 5.5 Flash arrives for everyone who just wants to use it.
If you are a researcher, a domain-model builder or a coding-agent team inside the licence territory with access to server GPUs, ERNIE 5.5 Flash Base is worth evaluating this week — the weights are free, the serving recipes are documented, and the efficiency story is the most interesting part. If you are anyone else, the right move is to watch: watch for independent benchmarks, watch for community ports, and watch for the post-trained release. The base model is the foundation. The finished model is the one most people will actually use, and it has not shipped yet.
All benchmark figures in this article are Baidu’s own, as published on the official model card, and have not been independently verified. Licence terms are summarised from the ERNIE 5.5 Flash Base Model License Agreement v1.0; read the full agreement on Hugging Face before building a product on the model.
