What Baidu released

On 5 November Baidu published ERNIE 5.5 Flash Base, a text-only sparse mixture-of-experts (MoE) model. An MoE model holds many small "expert" sub-networks and sends each token through only a few of them. That keeps the compute per token far below what the total parameter count suggests. Baidu's model card lists 280B total parameters with 14B active per token. The model has 43 layers, 512 routed experts per MoE layer (10 are picked per token) plus one shared expert, and was pre-trained on 33.5 trillion tokens. Baidu released FP8 weights in the main repository and a BF16 copy separately.

The checkpoint is a base model. Baidu's release page says the weights come from the end of pre-training, before any mid- or post-training. The model continues text and won't follow instructions the way a chat model does. Baidu frames the release as a shared starting point for domain adaptation, instruction tuning, agent training and reinforcement learning research. The same page says a post-trained ERNIE 5.5 Flash "will follow", with no date given. Trending Topics reports that this is Baidu's first large open-weight model since ERNIE 4.5 in mid-2025. ERNIE 4.5 shipped under Apache 2.0.

How the architecture saves memory

Most of the engineering goes into making long contexts cheaper. The model card shows that 37 of the 43 layers use Heavily Compressed Attention, which stores one key-value cache entry per 128 tokens. The other six layers use Multi-head Latent Attention with sparse attention and keep access to the full context. The key-value (KV) cache is the memory a server holds for every token it has already processed. Compressing most layers this hard is what makes Baidu's claimed 1,048,576-token inference context plausible on a fixed GPU budget. Training ran at 131,072 tokens. The model card lists the million-token figure for fine-tuning and inference.

The model also uses a technique Baidu calls ECHO. ECHO reuses the final layer to predict upcoming tokens, so that layer can act as a built-in draft model for speculative decoding. In speculative decoding a cheap model guesses several tokens ahead and the full model checks them in one pass. The model card says ECHO adds 0.36B parameters, and they aren't counted in the headline figures.

Benchmarks, with caveats

Baidu ran ERNIE 5.5 Flash Base and three competing base models (DeepSeek-V4-Flash-Base, Tencent's Hy3-preview-Base and Xiaomi's MiMo-V2.5-Base) through one evaluation pipeline. In the model card's comparison table ERNIE scores 78.4 on MMLU-Pro against DeepSeek's 69.2, and 82.3 on HumanEval+ against 69.5. It scores 91.7 on the RepoQA long-context code test, just ahead of Hy3's 91.3. ERNIE trails on six of the 30 tests, including HELMET-Rerank, where DeepSeek scores 60.0 to ERNIE's 56.2.

Every number in the table comes from Baidu. MarkTechPost points out that Baidu's pipeline gives rivals slightly different scores from the ones on their own model cards. We haven't found an independent evaluation yet. Benchmarks on a base model also tell you little about how a fine-tuned version will behave in production.

What it takes to run

This won't fit on a workstation. The FP8 checkpoint alone is about 280 GB of weights, and Trending Topics puts the requirement at at least 288 GB of GPU memory before you budget for KV cache. Baidu doesn't publish a minimum GPU spec. Both serving recipes on the model card use four-way parallelism. The SGLang recipe runs a dedicated Docker image (lmsysorg/sglang:dev-ernie-55-flash) with --tp 4. The vLLM path needs vLLM 0.32.0 plus Baidu's vLLM-ERNIE plugin. The model card shows this vLLM launch:

vllm serve baidu/ERNIE-5.5-Flash-Base \
  --kv-cache-dtype fp8 \
  --enable-expert-parallel \
  --data-parallel-size 4 \
  --max-model-len 131072 \
  --port 30000 \
  --moe-backend humming

Because it's a base model, you send raw prompts to the /v1/completions endpoint and the model carries on from where your text stops. Coverage at launch from eesel AI said no inference provider was serving the model yet and Baidu hadn't published a per-token price. For now, trying it means renting a multi-GPU node.

Read the licence first

The weights come under a custom ERNIE 5.5 Flash Base Model License Agreement v1.0. For most builders its terms matter more than the benchmarks. We've summarised the main clauses below. Our summary isn't legal advice.

  • Territory (§1.11): the licence covers "all geographical areas worldwide other than the European Union, the United Kingdom and South Korea." New Zealand is inside the territory. Trending Topics reads the restriction as applying to where the model is used, wherever the company is based. Outside the territory, nobody may use or display the model or its outputs.
  • Non-commercial use (§4.1): free, with no notice required.
  • Commercial use up to US$50 million in annual revenue from the related product (§4.2): you must email Baidu a one-time notice before first use. You don't need to wait for a reply.
  • Commercial use above US$50 million (§4.3): needs separate written authorisation from Baidu.
  • Attribution (§3.2): commercial products must display "Powered by ERNIE" prominently.
  • Training restriction (§5.5): you can't use the model or its outputs to improve any AI system other than ERNIE and its derivatives. You can fine-tune, quantise and merge ERNIE itself.
  • Military use: banned under Exhibit A.
  • Governing law (§9): Hong Kong law, with disputes heard in Hong Kong courts.

For an Auckland team the territory clause is the practical risk. A product with users in the UK or EU can't show ERNIE outputs to them, so anything sold into those markets needs geo-gating or a different model. Section 5.5 also rules out a common workflow, where a large model generates synthetic data to train a smaller in-house model. That's only allowed if the smaller model is itself an ERNIE derivative. MarkTechPost's comparison notes that several open-weight base models of similar size, including DeepSeek-V4-Flash and MiMo-V2.5, use the MIT licence.

What a builder could do with it

The release suits research groups and teams that run their own post-training. A university lab or a startup testing instruction tuning or RL recipes gets a strong, well-documented base checkpoint, free for non-commercial work. A long-context code model with a reported RepoQA score of 91.7 could also be a starting point for a fine-tuned repository assistant. Anyone going that way should test the self-reported long-context numbers on their own code first and confirm that every customer sits inside the licensed territory.

If you just want a capable model to call, wait for the post-trained ERNIE 5.5 Flash or an independent evaluation. If you fine-tune or benchmark the base model, we'd like to see your results on the ASAI projects page.

Sources

Build with people working on the same things

Join the Auckland Society for AI: NZ$599 a year with a ForgeBox and the members' projects area. Our Auckland meetups are open to everyone.

Become a member