Baidu released ERNIE 5.5 Flash Preview on 11 November 2026, the first post-trained version of the ERNIE 5.5 Flash Base checkpoint it published on 5 November. It has 280B total parameters, 14B active, and a 1M-token context (Baidu ERNIE, 2026). Short answer for local users: not yet. Today it runs only through Baidu's API, and the weights are promised for later this month.
TL;DR: ERNIE 5.5 Flash Preview is API only for now, on Baidu Qianfan and free on OpenRouter. Baidu says the weights will be open-sourced before 25 November under Apache 2.0. Going by the Base checkpoint it is built on, the FP8 weights should take about 288 GB, which fits a 512GB Mac Studio on paper, but neither llama.cpp nor mlx-lm supports the architecture yet.
What did Baidu confirm?
Two official sources: Baidu's release post, in Chinese and English, and the model's OpenRouter listing, whose endpoints API gives the output limit.
| Spec | Value |
|---|---|
| Total parameters | 280B |
| Activated parameters | 14B |
| Context length | 1M tokens (OpenRouter lists 1,024,000) |
| Max output on OpenRouter | 921,600 tokens |
| Input and output | Text only (OpenRouter) |
| Base model | ERNIE 5.5 Flash Base |
| Access today | Baidu Qianfan API, free on OpenRouter |
| Open weights | Before 25 November, Apache 2.0 (EN page) |
Get told when a better model fits your MacBook Air M5 16 GB
One email when a new open-weight model is a better fit for your machine, plus the Thursday weekly.
For: MacBook Air M5 16 GB
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
The Chinese page says the weights "will be open-sourced before November 25" (权重将于11月25日前开源). The English page says they "will be open-sourced under Apache 2.0 soon". If that licence holds, it is a change from the Base checkpoint, which ships under Baidu's own ERNIE 5.5 Flash Base Model License with revenue and territory limits.
Baidu describes the model as strong at coding, 3D modeling, knowledge work and content creation, and calls its decoding fast. Its page has a benchmark table in image form; we have not reproduced it, so we do not quote it. We found no tokens-per-second figure on either official page, and OpenRouter shows no throughput figure for the endpoint yet, so this article gives no speed.
What is still missing?
- No weights. The Baidu organisation on Hugging Face lists
ERNIE-5.5-Flash-BaseandERNIE-5.5-Flash-Base-BF16, but no Preview repository as of 12 November 2026, per the Hugging Face API. - No llama.cpp support. A pull request adding ERNIE 5.5, llama.cpp #31925, was opened on 9 November, is still a draft and is not merged. The master branch has entries for ERNIE 4.5 only.
- No mlx-lm support. mlx_lm/models has
ernie4_5.pyandernie4_5_moe.py, but nothing for ERNIE 5.5. - No model card. The architecture, size on disk and final licence text will only be certain once Baidu publishes the repository.
Can a 512GB Mac Studio run ERNIE 5.5 Flash Preview?
Not today, because there are no weights to load. Memory should not be the problem once they ship. Baidu says the Preview is a post-trained version of the Base, with the same 280B and 14B counts, so the closest real file sizes are the Base checkpoint's: 288,416,689,496 bytes in FP8 per its safetensors index, and 572,926,369,752 bytes in BF16 per the BF16 index. Neither has been published for the Preview.
| Build | Size | Status |
|---|---|---|
| FP8 (Base) | about 288 GB | Base only |
| BF16 (Base) | about 573 GB | Base only, over 512 GB |
| 4-bit (est.) | about 158 to 168 GB | Estimate, no file |
The 4-bit line is our estimate, not a file: 280B parameters at 4.5 to 4.8 bits per weight, the range typical of llama.cpp's Q4_K_M, before any context cache. The largest Mac Studio has 512GB of unified memory, per Apple's tech specs. On that machine an FP8 build would take a little over half the memory, and a 4-bit build about a third, leaving room for a long context cache. These figures assume the Preview keeps the Base's size, which Baidu's parameter counts suggest but its repository will confirm.
So the distance to a 512GB Mac Studio is not memory. It is two missing pieces: the weights, promised before 25 November, and a Mac engine that can load the new architecture.
What should you watch for?
1. The Hugging Face repository. Its config.json and safetensors index will confirm the architecture and the real size.
2. The licence file. Apache 2.0 is on Baidu's English page; the repository's LICENSE is what counts.
3. llama.cpp #31925 or an mlx-lm pull request. Without one of them merged, no GGUF or MLX build will run on a Mac.
ModelFit will add a model page and a dataset entry once the weights are public and llama.cpp or mlx-lm supports them.
What can you run instead today?
A model in the same size class that already works locally. DeepSeek V4 Flash runs on a Mac today, and our DeepSeek V4.1 Flash memory guide gives the RAM needed per build. For another big-lab preview that is cloud only, with no weights promised, see Qwen4-Max Preview. For a pick matched to your hardware, use the ModelFit wizard or npx @modelfit/cli@latest.
FAQ
Can you run ERNIE 5.5 Flash Preview locally?
Not yet. It is available only through Baidu Qianfan and OpenRouter. Baidu says the weights will be open-sourced before 25 November 2026.
How big is ERNIE 5.5 Flash Preview?
280B total parameters with 14B active, per Baidu. The Base checkpoint it is built on takes about 288 GB in FP8.
What licence will ERNIE 5.5 Flash Preview use?
Baidu's English release page says Apache 2.0. The Base checkpoint uses Baidu's own licence, so check the LICENSE file when the repository goes up.
Does llama.cpp support ERNIE 5.5?
Not yet. Pull request #31925 adds it but is an unmerged draft as of 12 November 2026. mlx-lm has no ERNIE 5.5 support either.
Is ERNIE 5.5 Flash Preview free?
On OpenRouter, the baidu/ernie-5.5-flash-preview:free endpoint is listed at $0 per token for prompts and completions.
Sources
- Baidu ERNIE, ERNIE 5.5 Flash Preview release (Chinese), 11 November 2026: https://ernie.baidu.com/zh/release/ernie-5-5-flash-preview/
- Baidu ERNIE, ERNIE 5.5 Flash Preview release (English): https://ernie.baidu.com/release/ernie-5-5-flash-preview/
- OpenRouter, ERNIE-5.5-Flash-Preview (free): https://openrouter.ai/baidu/ernie-5.5-flash-preview:free
- OpenRouter endpoints API, read 12 November 2026: https://openrouter.ai/api/v1/models/baidu/ernie-5.5-flash-preview:free/endpoints
- Hugging Face API, Baidu ERNIE 5.5 models, read 12 November 2026: https://huggingface.co/api/models?author=baidu&search=ERNIE-5.5
- ERNIE 5.5 Flash Base safetensors index (FP8): https://huggingface.co/baidu/ERNIE-5.5-Flash-Base/resolve/main/model.safetensors.index.json
- ERNIE 5.5 Flash Base safetensors index (BF16): https://huggingface.co/baidu/ERNIE-5.5-Flash-Base-BF16/raw/main/model.safetensors.index.json
- llama.cpp pull request #31925, model: add ERNIE 5.5: https://github.com/ggml-org/llama.cpp/pull/31925
- mlx-lm models directory: https://github.com/ml-explore/mlx-lm/tree/main/mlx_lm/models
- Apple, Mac Studio tech specs: https://www.apple.com/mac-studio/specs/
Don't miss the next one for your MacBook Air M5 16 GB
This article covers one moment. Get one email when a newer open-weight model fits your machine better, plus the Thursday weekly with what changed.
For: MacBook Air M5 16 GB
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
Match this model to a machine that can run it: by RAM tier for Apple Silicon, or by VRAM for an NVIDIA GPU.
See how this changes your recommendation
Run the wizardHave questions? Reach out on X/Twitter