DEV Community

Cover image for What Is ERNIE 5.5 Flash?
Hassann
Hassann

Posted on Originally published at apidog.com

What Is ERNIE 5.5 Flash?

What is ERNIE 5.5 Flash? It’s Baidu’s newest open-weight model family: a sparse mixture-of-experts (MoE) language model with 280B total parameters, 14B active per token and a 1M-token context window. It ships as two separate artifacts. ERNIE 5.5 Flash Base (November 5, 2026) is a raw pretraining checkpoint on Hugging Face. ERNIE 5.5 Flash Preview (November 11, 2026) is the post-trained chat and agent model, live as an API on Baidu Qianfan and as a free endpoint on OpenRouter.

Try Apidog today

What is ERNIE 5.5 Flash?

ERNIE 5.5 Flash is a 280B-parameter sparse MoE language model from Baidu that activates 14B parameters per token and supports a 1M-token context. It’s text-only, trained from scratch on 33.5T tokens, and built for coding, agentic tasks, knowledge work and long documents. It comes as a Base checkpoint and a post-trained Preview.

ERNIE 5.5 Flash

Baidu’s pitch is efficiency. The Base release page says every component was redesigned to balance capability against activated compute, KV cache size and long-context cost. The Preview page calls the result “strong at coding, 3D modeling, knowledge work and content creation.”

It succeeds ERNIE 5.1; for background, see What is ERNIE 5.1 and the ERNIE 5.1 API guide.

ERNIE 5.5 Flash Preview vs Base

The naming trips people up, so here’s the side-by-side:

ERNIE 5.5 Flash Base ERNIE 5.5 Flash Preview
Released November 5, 2026 November 11, 2026
What it is End-of-pretraining checkpoint, no mid- or post-training First post-trained model built on Base, an intermediate checkpoint
Use it for Fine-tuning, domain adaptation, RL research Chat, coding agents, tool calling
How to call it Self-host; completions endpoint only Qianfan API, OpenRouter baidu/ernie-5.5-flash-preview:free
Weights Public on Hugging Face (FP8 and BF16) Not public yet; Baidu says Apache 2.0 “soon”
Licence ERNIE 5.5 Flash Base Model License v1.0 Apache 2.0 promised, not yet published
Size 280B total, 14B active 280B total, 14B active
Context 128K training, 1M for fine-tuning and inference 1,024,000 tokens on OpenRouter

Diagram comparing ERNIE 5.5 Flash Base, a pretraining checkpoint on Hugging Face, with ERNIE 5.5 Flash Preview, the post-trained chat model on Qianfan and OpenRouter, both sharing 280B total and 14B active parameters

The rule: to chat with it or plug it into a coding agent, use Preview. To train your own model on a strong pretrained start, use Base. Base isn’t instruction-tuned, so it won’t follow chat prompts.

Baidu is open about Preview being unfinished. The release page calls it “an intermediate checkpoint from post-training” and says “some capabilities may still fall short” until the official release.

ERNIE 5.5 Flash architecture

Here are the parts of Baidu’s technical write-up that matter to you as a user.

Hybrid sparse attention

The model has 43 Transformer layers. Only 6 of them (roughly one in seven) keep token-level access to the full context, using Multi-head Latent Attention (MLA) with sparse attention (DSA). The other 37 layers use Heavily Compressed Attention (HCA), which squeezes every 128 tokens into a single cached entry and attends over those plus a short window of recent tokens.

Why you care: the KV cache stays small, so a 1M-token context is cheaper to serve than the parameter count suggests. Baidu says the sparse attention in the MLA layers keeps the per-query cost of core attention constant as context grows.

Multi-Router latent MoE

A normal MoE layer uses one router to pick its top-k experts. ERNIE 5.5 Flash uses several independently initialized routers that score experts jointly, which Baidu says corrects bad expert choices that a single router would reinforce. Each MoE layer holds 512 routed experts (10 active per token) plus 1 shared expert, running in a 2048-dimension latent space to roughly halve expert compute and traffic.

Other changes

  • MultiMax output head: a learnable piecewise softmax temperature.
  • ECHO instead of MTP: the final layer predicts future tokens and doubles as a speculative-decoding draft model.
  • Tokenizer: about 200K entries, with numbers split right to left in groups of three digits.

Architecture diagram of ERNIE 5.5 Flash showing 43 layers, 37 compressed HCA attention layers, 6 full-access MLA layers, and a Multi-Router MoE block with 512 experts and 10 active per token

ERNIE 5.5 Flash benchmarks

All numbers below are Baidu’s own, from its release pages. Many comparison scores carry Baidu’s asterisk, meaning Baidu ran the evaluation itself rather than quoting the other lab.

Preview (post-trained), selected rows from Baidu’s table:

Benchmark ERNIE 5.5 Flash Preview DeepSeek-V4.1-Flash GPT-6.1 Sol Claude Opus 5.5
Terminal-Bench 4.0 46.5 42.3 56.1 66.4
DeepSWE v1.1 71.2 77.0 75.2 74.2
MCP-Atlas 84.6 77.8 81.0 84.2
AutomationBench 61.9 60.5 49.3 58.0
BrowseComp 87.4 80.3 87.5 90.8
HLE (no tools, text-only) 46.4 44.0 52.9 61.4

Read that honestly: Preview beats its size class on several agent and knowledge-work tests (MCP-Atlas, AutomationBench, SpreadsheetBench) and trails frontier models on hard agentic coding like Terminal-Bench. It lands below DeepSeek-V4.1-Flash on DeepSWE. For a full head-to-head with DeepSeek, see the sibling post ERNIE 5.5 Flash vs DeepSeek V4.1 Flash, and our DeepSeek V4.1 Flash explainer for the other side.

Base (pretrained), against base models of similar size: Baidu reports MMLU-Pro 78.4 vs 69.2 for DeepSeek-V4-Flash-Base, HumanEval+ 82.3 vs 69.5, LiveCodeBench 40.1 vs 31.9 and LongBench-v2 48.7 vs 44.5. It loses a few rows, such as InfiniteBench. The full table is on the Hugging Face model card.

On speed, OpenRouter’s description says “more than 200 tokens per second.” OpenRouter’s own 3-day dashboard on November 12 showed a P50 throughput of 79 tokens per second on the free endpoint, which is shared traffic, so expect your numbers to vary.

How to access ERNIE 5.5 Flash

You have three routes today:

  1. OpenRouter (free, Preview): call baidu/ernie-5.5-flash-preview:free through OpenRouter’s OpenAI-compatible API. Baidu Qianfan is the only provider, and the model page lists $0 input and output. Free endpoints are rate limited per OpenRouter’s limits docs. Full walkthrough in the sibling ERNIE 5.5 Flash free guide.
  2. Baidu Qianfan (Preview): the official API, linked from Baidu’s release page. See the ERNIE 5.5 Flash API guide for request setup.
  3. Hugging Face (Base only): download FP8 or BF16 weights and serve them with vLLM (plus the vLLM-ERNIE plugin) or SGLang. Baidu’s examples use 4-way parallelism on a multi-GPU node. Step-by-step in run ERNIE 5.5 Flash locally with vLLM.

A minimal OpenRouter call looks like this:

curl https://openrouter.ai/api/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $OPENROUTER_API_KEY" \
  -d '{
    "model": "baidu/ernie-5.5-flash-preview:free",
    "messages": [{"role": "user", "content": "Explain MoE routing in two sentences."}]
  }'
Enter fullscreen mode Exit fullscreen mode

OpenRouter says the endpoint accepts tools, tool_choice, response_format with a JSON schema and a reasoning parameter. If you’d rather drive it from a terminal agent, the sibling post on ERNIE 5.5 Flash in OpenCode for free covers that setup.

ERNIE 5.5 Flash licence: what you can and can’t do

The two artifacts have different terms, and the difference matters if you plan to ship a product.

Base uses the ERNIE 5.5 Flash Base Model License Agreement v1.0, not Apache 2.0. In short:

  • Non-commercial use (research, education, evaluation, personal experiments) is free with no notice.
  • Commercial use is free when the annual revenue of the related products is at or below US$50 million, provided you send Baidu a one-time use notice and display “Powered by ERNIE”.
  • Above US$50 million, you need separate written authorization from Baidu before you start.

Preview has no published weights or licence file yet. Baidu says the weights “will be open-sourced under Apache 2.0 soon” but gives no date. Until that file exists, treat Preview as API-only.

Who should use ERNIE 5.5 Flash?

  • Teams running agents on a budget. A free OpenRouter endpoint with 1M context and tool calling is a cheap way to test agent loops.
  • Fine-tuning and RL researchers. Base is a clean end-of-pretraining checkpoint, which is rare at this size.
  • Long-document workloads. The compressed attention design targets 1M-token inputs without a huge KV cache.

Where to hold back:

  • Strict JSON pipelines. OpenRouter’s dashboard showed a 21.70% structured-output error rate for the free endpoint on November 12. Validate every response.
  • Image input. The model is text-only; Baidu’s own benchmark notes say image understanding was supplied by an external tool.
  • Production SLAs. Preview is an unfinished checkpoint served by a single provider.

Let your coding agent test ERNIE 5.5 Flash with Apidog CLI

That 21.70% structured-output error rate is why you test the response shape before building on it. Let Claude Code, Codex or Cursor create the endpoint and test case with Apidog CLI.

Set it up once:

npm install -g apidog-cli@latest
apidog auth login --with-token <YOUR_APIDOG_TOKEN>
apidog skill install
Enter fullscreen mode Exit fullscreen mode

Then give your agent one prompt:

In Apidog project <projectId>, create an environment "openrouter" with base URL
https://openrouter.ai/api/v1 and my OpenRouter key as a variable. Add a POST
/chat/completions endpoint for model baidu/ernie-5.5-flash-preview:free.
Create a test case that asserts HTTP 200, choices[0].message.content is a
non-empty string, and usage.total_tokens is a number. Run it and summarize failures.
Enter fullscreen mode Exit fullscreen mode

The agent checks each payload against the CLI schema before it writes anything:

apidog environment create openrouter --project <projectId> --base-url https://openrouter.ai/api/v1
apidog cli-schema get endpoint-create
apidog cli-schema validate endpoint-create --file endpoint.json
apidog endpoint create --project <projectId> --file endpoint.json
apidog test-case category --project <projectId>
apidog cli-schema validate test-case-create --file case.json
apidog test-case create --project <projectId> --file case.json
apidog run --test-case <caseId> --project <projectId> -e <envId> -r cli,json
Enter fullscreen mode Exit fullscreen mode

Swap the base URL for your Qianfan or local vLLM endpoint and the same case checks those too. To run it on every commit, add -r junit and follow the Apidog CLI CI/CD guide. New to the CLI? Start with Apidog CLI in Claude Code. Apidog Free covers up to 4 users, so download Apidog and try it on your own project.

FAQ

Is ERNIE 5.5 Flash open source? Partly. Base weights are public under Baidu’s own licence, free for commercial use up to US$50 million in related annual revenue. Preview weights are promised under Apache 2.0 “soon”.

What’s the difference between ERNIE 5.5 Flash Preview and Base? Base is the raw pretraining checkpoint for fine-tuning. Preview is the post-trained chat and agent model you call through an API. Both have 280B total and 14B active parameters.

Is ERNIE 5.5 Flash free? The Preview is free on OpenRouter as baidu/ernie-5.5-flash-preview:free, with rate limits. Base is free to download. Baidu hasn’t published a paid per-token price on an official page as of November 12, 2026.

How big is the ERNIE 5.5 Flash context window? 1M tokens: 1,024,000 for Preview on OpenRouter and 1,048,576 in the Base config, though Base trained at up to 128K.

Can I run ERNIE 5.5 Flash locally? Base, yes, with vLLM plus the vLLM-ERNIE plugin or SGLang on a multi-GPU node. Preview can’t run locally until Baidu publishes its weights.

Does ERNIE 5.5 Flash support images? No. Both versions are text-only models.

Top comments (0)