Baidu releases ERNIE 5.5 Flash Base: the flash model race extends to China

Baidu releases ERNIE 5.5 Flash Base: the flash model race spreads to China

🔎 A Chinese giant enters the war of small, fast models

On November 5, 2026, Baidu published the weights of ERNIE 5.5 Flash Base on Hugging Face, a mixture-of-experts (MoE) language model with 280 billion total parameters, of which only 14 billion are active per token. A format that immediately brings to mind the wave of "flash" models that has swept the market since the summer: GLM-5.3 Flash from Z.ai, Qwen3.8 Flash from Alibaba, MiniMax M3.1 Flash Preview.

Why now? Because the niche of open, fast, and cheap models has become the most profitable battleground in AI in 2026. Chinese labs have realized that the majority of production queries — classification, extraction, agent routing, summarization — don't require a frontier model scoring 90+, but rather a fast model costing a few fractions of a cent per million tokens. After GLM-5.3-Flash and Qwen3.8-Flash-Next: China releases two Flash models on the same day — the price war shifts into high gear, it's now Baidu's turn to play.

But there's a subtlety that changes everything: ERNIE 5.5 Flash Base is an end-of-pretraining checkpoint, with no post-training and no chat template. It's not a chatbot. It's a research artifact intended for developers who want to do their own fine-tuning. A deliberate strategic choice, confirmed by Baidu's technical blog, which specifies that the post-trained model will "follow".

The Essentials

• ERNIE 5.5 Flash Base is a text-only MoE with 280B parameters / 14B active per token, released as open weights on November 5, 2026, with a 128K context in training and up to 1M tokens in inference and fine-tuning.
• It's a base model: an end-of-pretraining checkpoint, without instruction tuning. FP8 and BF16 weights are available in safetensors on Hugging Face and ModelScope, under a bespoke license (ERNIE 5.5 Flash Base License v1.0).
• Baidu claims scores on 30 benchmarks against DeepSeek-V4-Flash-Base, Hy3-preview-Base and MiMo-V2.5-Base — but no independent benchmark exists yet, and no Artificial Analysis page either.
• Baidu is entering a low-cost niche that is already very crowded (Qwen3.8 Flash, GLM-5.3 FlashX, MiniMax M3.1 Flash) and without a published price to date.
• The license is tighter compared to ERNIE 4.5's Apache 2.0 in 2025 — a step backward for the open source community.

Recommended tools

If you want to run this type of model on your own machine, our guide on installing a local LLM with Ollama or LM Studio covers the complete steps, and our comparison of the best LLMs to run locally will help you choose based on your hardware.

ERNIE 5.5 Flash Base: the specs, with numbers to back them up

Direct answer: it's a 280B-parameter MoE with 14B active per token, trained on 33.5 trillion tokens, released in open weights on November 5, 2026.

Let's break down the figures, sourced from OrcaRouter's detailed analysis:

Two points deserve a closer look.

The 280B/14B ratio is aggressive. With only 5% of parameters active per token, Baidu is clearly targeting latency and inference cost. It's the same bet as DeepSeek-V4-Flash-Base (284B-A13B) and MiMo-V2.5-Base (310B-A15B), its two direct competitors cited in the benchmarks. The economic logic is simple: a low-activation MoE costs nearly the price of a dense 14B model, while retaining the knowledge capacity of a 280B model.

1M context in inference is a signal for agents. A million tokens of context opens the door to agent workflows that need to digest entire codebases or document corpora. It's the same argument Cloudflare makes with Clef and Clef-flash: decision models for agents, not chatbots — the flash model market is no longer about chat, but about agentic infrastructure.

Why a "Base" model and not a chat model?

Direct answer: a base model is a raw checkpoint from the end of pretraining, without conversational alignment. It doesn't respond to instructions naturally — it completes text. It's a choice for developers, not for end users.

The official Hugging Face card and Baidu's blog are explicit: no post-training, no chat template. The post-trained model "will follow."

Concretely, what does this change for you?

If you're a developer who fine-tunes

This is excellent news. A clean base model, with FP8 weights (for economical inference) and BF16 (for training), gives you a quality blank canvas. You can:

• Do continued pretraining on your domain (legal, medical, finance).
• Apply SFT then DPO/RLHF to create your own instruct model.
• Use fine-tuning on 1M context — rare at this model tier, and valuable for long-document tasks.

If you just want a chatbot

Move along, for now. Without a chat template, plugging ERNIE 5.5 Flash Base into a conversational pipeline will produce erratic results. Wait for the post-trained model Baidu promises, or look at already-aligned models instead. Our comparison of the best free LLMs (ChatGPT free, Gemini, Groq) covers the options you can use right away.

If you're building agents

It's more nuanced. Modern agents often work better on base models fine-tuned for specific tasks (routing, tool selection, decision-making) than on over-prompted generalist chatbots. This is exactly Cloudflare's thesis with Clef, and it's probably Baidu's real target here. Our feature on the best LLMs for AI agents details this paradigm.

The flash model war: where does Baidu stand?

Short answer: Baidu is arriving late to a market that Z.ai, Alibaba and MiniMax have already saturated, with a bigger model but no published pricing or endpoint.

The timeline speaks for itself:

Three ways to read this table.

First reading: Baidu is the only one releasing a base model. GLM-5.3 Flash is multimodal, under MIT license, with a 1M context window — a finished product. Qwen3.8 Flash is a mature API offering. ERNIE 5.5 Flash Base, on the other hand, is a research artifact. It's a different positioning: Baidu isn't selling yet, it's inviting the community to build. It's a strategy that has proven itself for Alibaba with the Qwen family, but it's arriving several quarters late.

Second reading: the price war is in full swing. According to Shattered.io's market data, API prices in November 2026 are dizzying:

(Third-party listing prices, November 2026 — check the providers' websites, they change every week.)

At this level, every fraction of a cent counts. ERNIE 5.1, the previous generation, was positioned around $0.59/$0.65 per million tokens on international Qianfan. If Baidu wants to be competitive in the flash segment, it will have to do better than Qwen3.8 Flash at $0.15/$0.47. There's no indication yet that it will — the price simply isn't published.

Third reading: volume follows speed. We observed this with MiniMax M3.1 Flash Preview: China continues its war of fast small models — every Chinese lab is now releasing its flash iteration every one to two months. ERNIE 5.5 Flash Base fits into this industrial cadence. It's not an isolated launch, it's a position in a race.

The black hole of independent benchmarks

Direct answer: the only figures available on ERNIE 5.5 Flash Base come from Baidu itself, across 30 benchmarks against three competing models. No independent evaluation exists yet.

This is the most important point of this entire article, and I'll be direct: beware of vendor-reported scores on a base model.

Here's what we know. Baidu compares its model to:

• DeepSeek-V4-Flash-Base (284B-A13B)
• Hy3-preview-Base (295B-A21B) — Tencent's line of open models, whose release we covered in Tencent opens Hy4: 770B parameters, native vision and computer use in open weight
• MiMo-V2.5-Base (310B-A15B)

Three critical observations:

1. They're all base models. Comparing pretraining checkpoints against each other is legitimate in research, but these scores don't predict the production performance of a derived instruct model.
2. There is no Artificial Analysis page for ERNIE 5.5 Flash Base to date. No independent replication. The scores are self-reported, period.
3. The choice of competitors is telling. Baidu compares against similarly sized models (280-310B), never against the smaller, faster flash models that dominate the price/performance market. That's favorable framing.

For context: today's frontier models like Gemini 3.1 Pro (92), GPT-5.5 (91), or Claude Opus 4.7 Adaptive (90) sit in a different performance category — and price bracket. ERNIE 5.5 Flash Base isn't competing with them, and doesn't claim to. The real question is: once fine-tuned, will this 280B-A14B beat Qwen3.8 Flash on high-volume tasks? No one can say today.

Our advice: wait for independent evaluations, or run your own on your own dataset. That's actually the whole point of a base model — you can evaluate it precisely on your own use case before investing.

The bespoke license: a step back from ERNIE 4.5

Direct answer: yes, Baidu is tightening the screws. ERNIE 4.5 was under Apache 2.0 in 2025; ERNIE 5.5 Flash Base arrives under a bespoke license, the ERNIE 5.5 Flash Base License v1.0.

This is an important signal for the open source community, and it needs to be read against what the competition is doing:

• Z.ai releases GLM-5.3 Flash under MIT — the most permissive license in existence. Commercial use, modification, redistribution, everything is permitted.
• Alibaba keeps the Qwen family mostly under Apache 2.0.
• Baidu chooses a custom proprietary license, whose exact commercial restriction terms still need to be audited case by case.

Why this step back? Probably a combination of factors: protecting the commercial value of the upcoming instruct model, fear of distillation by competitors, and internal pressure to monetize Qianfan. It's the same dilemma the whole industry is going through: Meta eventually loosened its Llama license, Mistral oscillates between open weights and closed models — Mistral Large 4 was, incidentally, released with open weights on October 29, 2026, proof that the movement doesn't only go one way.

For a developer, the practical consequence is simple: read the license before building a product on top of it. A bespoke license can prohibit certain commercial uses, impose user thresholds, or require negotiation beyond a certain volume. It's not a dealbreaker, but it's a friction that MIT and Apache 2.0 don't have.

What This Says About the Market: China Sets the Pace for Low-Cost

Direct answer: in 2026, price/performance innovation on open models is coming overwhelmingly from China, and Western labs are following rather than leading on this segment.

Look at the trajectory over twelve months. Summer 2026 saw Z.ai and Alibaba release flash models almost simultaneously. Fall saw MiniMax keep the streak going, Mistral respond with Mistral Large 4 in open weights, Cloudflare enter the fray with its Clef decision models. And now Baidu, the last major Chinese lab that hadn't yet played its flash card, is playing it.

The economic implications are considerable:

The marginal cost of AI is collapsing on high-volume tasks. When a flash model costs $0.15 per million input tokens, classifying a million support tickets costs $150. In 2024, the same operation with a frontier model cost dozens of times more. This makes economically viable use cases that weren't before: document analysis at scale, massive moderation, structured extraction across entire corpora.

The model hierarchy is stratifying. The market is clearly segmenting: frontier models (Gemini 3.1 Pro, GPT-5.5, Claude Opus 4.7) for complex tasks and heavy reasoning; mid-tier models (Claude Sonnet 4.6, GLM-5.1) for general-purpose work; flash models for volume. Our monthly comparison of the best LLMs reflects this stratification, and our feature on the best LLMs for research shows how Perplexity and NotebookLM are already leveraging this hierarchy.

Base models are becoming a product in their own right. This is perhaps the most interesting trend. Publishing a raw pretraining checkpoint is betting on the fine-tuning ecosystem — training frameworks, GPU providers, ML teams that prefer to control their own post-training. It's a way of saying: "we provide the raw material, you build." If this logic takes hold, we'll see more and more labs publishing base models as ecosystem bait.

How to actually test ERNIE 5.5 Flash Base

Direct answer: download the weights from Hugging Face or ModelScope, but be prepared for serious hardware and a fine-tuning pipeline — this is not a plug-and-play model.

The practical steps:

1. Download the weights in safetensors format from Hugging Face. Plan on the FP8 version for inference (lighter) and BF16 for training.
2. Watch out for the size. 280B parameters, even as MoE, won't fit on a laptop. In FP8, expect several hundred GB of VRAM for comfortable inference. Multi-GPU setups are practically mandatory.
3. No chat template: if you want to test it conversationally, you'll need to apply your own template or wait for the instruct model.
4. For fine-tuning on a 1M context, check that your training framework is compatible with the 512-expert MoE architecture — not all of them handle it efficiently.
5. A realistic alternative: wait for community quantizations and local runners. Our guide to setting up a local LLM will be updated once accessible quantized versions become available.

Let's be honest: for most developers, ERNIE 5.5 Flash Base today is a model to watch, not to deploy. Its real value will materialize with the post-trained model and the API pricing on Qianfan.

❌ Common Mistakes

Mistake 1: Confusing base model and chat model

This is the most frequent mistake with this kind of release. A base model completes text; it does not reliably follow instructions. Hooking it into a chatbot without post-training will produce incoherent outputs. The solution: wait for Baidu's instruct model, or plan your own SFT before any conversational use.

Mistake 2: Trusting vendor benchmarks without verification

Baidu's 30 benchmarks are self-reported, with no independent replication or Artificial Analysis page. The solution: build your own evaluation set on your real data before any commitment, and watch for independent replications in the coming weeks.

Mistake 3: Ignoring the bespoke license

Many developers see "open weight" and assume a permissive license. ERNIE 5.5 Flash Base License v1.0 is a custom proprietary license, very different from GLM-5.3 Flash's MIT or ERNIE 4.5's Apache 2.0. The solution: read the terms before any commercial use, and negotiate with Baidu if your volume exceeds the planned thresholds.

Mistake 4: Comparing prices too early

No API pricing has been published for ERNIE 5.5 Flash Base. Comparing it with Qwen3.8 Flash ($0.15/$0.47) or GLM-5.3-FlashX ($0.37/$0.25) is premature. The solution: wait for the pricing announcement on Qianfan, and check up-to-date prices on aggregators before making any architecture decisions.

❓ Frequently Asked Questions

Is ERNIE 5.5 Flash Base free?

The weights are freely downloadable from Hugging Face and ModelScope, under the bespoke ERNIE 5.5 Flash Base License v1.0. But "free" doesn't mean "unrestricted": the bespoke license may impose conditions on commercial use. And inference itself will cost money via Qianfan once Baidu publishes its pricing.

Can it compete with GPT-5.5 or Gemini 3.1 Pro?

No, and that's not its goal. Frontier models like GPT-5.5 (91) or Gemini 3.1 Pro (92) target complex reasoning at a premium price. ERNIE 5.5 Flash Base targets high volume at low cost: classification, extraction, agent routing. These are two different markets, and the comparison only makes sense for specific use cases.

When will the chat (instruct) model be released?

Baidu states on its technical blog that the post-trained model "will follow", without a specific date. Given the market's pace — Chinese labs iterate every one to two months — expect a release in the weeks following the base model's publication. Keep an eye on release aggregators so you don't miss it.

Should you prefer it over GLM-5.3 Flash?

Impossible to say today. GLM-5.3 Flash is a finished product: multimodal, MIT, 1M context, with published pricing. ERNIE 5.5 Flash Base is a research artifact with no chat template and no pricing. For immediate use, GLM-5.3 Flash is more mature. For custom fine-tuning, ERNIE could become interesting — provided the license works for you and independent benchmarks confirm Baidu's scores.

Why is Baidu releasing a base model rather than a chatbot?

Three likely reasons: to capture the fine-tuning community before competitors, to test pretraining quality without exposing themselves to criticism over alignment, and to prepare the commercial ground for the upcoming instruct model. It's also a way to generate visibility and ecosystem feedback at minimal cost.

What hardware do you need to run it?

With 280B parameters, even as a MoE with 14B active, you need serious multi-GPU infrastructure — several hundred GB of VRAM in FP8. This is not a model for an individual workstation. More aggressive community quantizations may lower that barrier, but at the cost of a quality degradation.

✅ Conclusion

ERNIE 5.5 Flash Base is more of a strategic move than a product: Baidu is entering the flash model war with a research checkpoint of 280B/14B active parameters, with no pricing and no chatbot, betting on the fine-tuning ecosystem — the real battle will begin with the instruct model and the pricing announcement. In the meantime, keep an eye on our monthly comparison of the best LLMs to track how the price/performance ratio evolves in this red-hot segment.