Baidu released ERNIE 5.5 Flash Preview on November 11, 2026. Baidu’s release page calls it the first post-trained model built on ERNIE 5.5 Flash Base, which Baidu put out earlier. It’s a mixture-of-experts model with 280B total parameters, 14B active per token and a 1M-token context, and Baidu pitches it at coding, 3D modeling, knowledge work and content creation.
You can call it today in two places: Baidu’s Qianfan platform and OpenRouter, where it’s listed as baidu/ernie-5.5-flash-preview:free at $0 for prompt and completion tokens. The weights aren’t out yet. Baidu says they “will be open-sourced under Apache 2.0 soon”, without a date on the release page.
All benchmark claims below come from Baidu. Baidu also says the Preview is an intermediate post-training checkpoint and that “some capabilities may still fall short” until the official release.
ERNIE 5.5 Flash Preview at a glance
ERNIE 5.5 Flash Preview (checked November 14, 2026)
| Released | November 11, 2026 (Baidu release page) |
|---|---|
| Parameters | 280B total, 14B active per token (MoE) |
| Context | 1M tokens per Baidu; OpenRouter lists 1,024,000 context and up to 921,600 completion tokens |
| Modality | Text only (text in, text out) |
| OpenRouter model ID | baidu/ernie-5.5-flash-preview:free |
| OpenRouter price | $0 prompt, $0 completion |
| Reasoning | Mandatory on OpenRouter; the only effort level listed is high |
| Tool use | OpenRouter lists tools, tool_choice, structured_outputs and response_format |
| Other API | Baidu Qianfan |
| Weights | Not released yet; Baidu says Apache 2.0, soon |
OpenRouter’s models API lists four endpoints for the model: /api/v1/chat/completions, /api/v1/completions, /api/v1/messages and /api/v1/responses.
What Baidu says it’s good at
Baidu’s release page is mostly a set of showcase projects. The model completed each one “end to end inside an agent harness”, with external assets credited separately:
- Game development. In a single 11-hour coding task, the model merged assets from several Unreal Engine generations into one Unreal Engine 5.8 project and wrote a third-person game entirely in C++. Baidu notes the model is text-only and can’t see the screen, so it turned playtest problems (a character hovering above the ground, a misplaced camera) into measurable tests in code.
- Mobile games from design docs. Working from complete game design documents, it built nearly 100 mobile mini-games with Cocos, Unity, LayaAir, PixiJS and Canvas 2D.
- 3D and CAD. Agents running the model used the Blender Python API to build an airport and an airliner, and wrote FreeCAD Python scripts for an editable W16 engine model.
- Knowledge work. Baidu shows generated research reports, a spreadsheet-based wealth planner and an orchestral piece delivered as WAV.
Baidu says it strengthened support for Unreal Engine, Unity and Godot with domain-specific training data. These are vendor demos, not independent tests, and the release page doesn’t say how many attempts each one took.
Try it free on OpenRouter
OpenRouter’s chat completions endpoint follows the OpenAI request format, so a quick test only needs an OpenRouter API key in OPENROUTER_API_KEY:
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "baidu/ernie-5.5-flash-preview:free",
"messages": [
{"role": "user", "content": "Write a Bash one-liner that lists the 10 largest files under /var/log."}
]
}'
Any OpenAI-compatible coding tool that lets you set a base URL of https://openrouter.ai/api/v1 and a model name can point at the same model ID. Because OpenRouter marks reasoning as mandatory for this model, expect reasoning tokens on every request, and don’t plan on turning reasoning off to save time.
Two practical points before you wire it into an agent:
- It’s a preview. Baidu says post-training is ongoing and the official release will bring improvements, so behaviour can change under the same model family.
- Free doesn’t mean private by default. Check OpenRouter’s provider and data policy settings for the account you use before you send proprietary code to any free model.
If you’re weighing OpenRouter against a vendor’s own API for coding tools, see OpenRouter vs direct API for Claude Code.
What’s inside: the ERNIE 5.5 Flash Base architecture
The Preview builds on ERNIE-5.5-Flash-Base, and its Hugging Face model card describes the architecture in detail:
- 280B-A14B hybrid-attention MoE with 43 layers: 1 dense FFN layer and 42 MoE layers.
- Experts: 512 routed experts (top-10) plus 1 shared expert per MoE layer.
- Attention: 37 Heavily Compressed Attention (HCA) layers, which cache a single entry per 128 tokens, and 6 Multi-head Latent Attention layers with sparse attention (MLA + DSA). Baidu says one in seven layers keeps full-context access through MLA.
- Training: pre-trained from scratch on 33.5T tokens with the Muon optimizer and a training context of up to 128K. Fine-tuning and inference support a 1M-token context.
- Weights: FP8 and BF16 safetensors, on Hugging Face and ModelScope.
That cache detail matters for long prompts. Only the 6 MLA layers keep a full-length cache, while the 37 HCA layers store one entry per 128 tokens, so the memory cost of a long context grows far more slowly than it would with full attention in every layer.
Baidu’s base-model table compares ERNIE-5.5-Flash-Base with other base models of similar size. A few of the code rows:
| Benchmark (Baidu-reported) | ERNIE-5.5-Flash-Base | DeepSeek-V4-Flash-Base | Hy3-preview-Base | MiMo-V2.5-Base |
|---|---|---|---|---|
| HumanEval+ (Pass@1) | 82.3 | 69.5 | 78.0 | 74.4 |
| MBPP+ (Pass@1) | 81.5 | 67.2 | 78.3 | 74.3 |
| LiveCodeBench (Pass@1) | 40.1 | 31.9 | 33.5 | 32.4 |
| FullStackBench (Pass@1) | 54.6 | 43.3 | 56.7 | 54.1 |
These are base-model scores from Baidu’s own evaluation pipeline. They don’t describe the post-trained Preview you call on OpenRouter.
Self-hosting: base model only, for now
You can’t self-host the Preview yet. The base model card documents SGLang and vLLM serving. Its vLLM example installs vllm==0.32.0 plus Baidu’s vLLM-ERNIE plugin and serves the model with four-way data parallelism:
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
That’s multi-GPU datacenter territory, not a single workstation card. As a base model it’s served through the completions endpoint, not chat, so it’s a starting point for fine-tuning and research rather than a drop-in coding assistant. For models you can run on one GPU, see Edge0-145B.
Before you adopt ERNIE 5.5 Flash Preview
- Test it on your own tasks; the showcase demos and benchmarks are Baidu's
- Use baidu/ernie-5.5-flash-preview:free on OpenRouter, or the Qianfan API
- Check OpenRouter data policy settings before sending private code
- Expect mandatory reasoning tokens on every request
- Don't assume the downloadable base weights are Apache 2.0; they use Baidu's own license
- Watch for the official release and the Apache 2.0 Preview weights
What isn’t known yet
A few things weren’t settled when we checked on November 14, 2026:
- Weight release date. Baidu’s English release page says “soon” and gives no date.
- A public Preview repo. OpenRouter’s listing points to
baidu/ERNIE-5.5-Flash-Previewas the Hugging Face ID, but that repo wasn’t publicly accessible when we checked. - Final scores. Baidu links a full benchmark table for the Preview, but it’s describing an intermediate checkpoint, and the official release may score differently.
Is it worth trying?
For a free, 1M-context model with tool calling, the barrier is low: one API key and one model ID. The open questions are how the Preview holds up outside Baidu’s showcases and when the promised Apache 2.0 weights land. If you’re already comparing open-weight coding models, add it to your test set alongside the options in our free AI coding tier roundup, and re-check the license and pricing once Baidu ships the official release.