There’s a new ERNIE checkpoint on Qianfan. Baidu calls it ERNIE 5.5 Flash Preview, and it’s the first post-trained model built on ERNIE 5.5 Flash Base, which the company put out on 5 November 2026. Total size is 280 billion parameters. Only 14 billion of those are active for each token. Context runs to 1 million tokens. Baidu’s English announcement says the weights will be open-sourced under Apache 2.0 soon. The Chinese announcement, and Baidu’s launch post, give a date: 25 November.
“ERNIE 5.5 Flash Preview is an intermediate checkpoint from post-training. It is already competitive with models of its size, but because training is not finished, some capabilities may still fall short.”
— Baidu, in the ERNIE 5.5 Flash Preview announcement
Once Baidu has watched how people use the Preview, it says it’ll ship an official release. The company is pitching the model for coding, 3D modelling, knowledge work and content creation. Its launch page is full of demos the model ran end to end in an agent harness, among them an 11-hour Unreal Engine 5.8 third-person game, plus Blender and CAD scenes, research reports and short videos. Baidu made those demos itself.
Baidu also published a benchmark appendix of its own. An asterisk on that chart marks a score Baidu measured itself. Unmarked scores come from the model’s official report or a public leaderboard. On Baidu’s chart, ERNIE 5.5 Flash Preview scores 61.9 on AutomationBench and 36.4 on SpreadsheetBench V2, the highest of the models listed there. On Terminal-Bench 4.0 it scores 46.5. GPT-6.1 Sol sits at 56.1, and Claude Opus 5.5 at 63.6, with the Claude figure marked as a self-evaluation. DeepSWE v1.1 has the Preview on 71.2, and Agents’ Last Exam (CLI) on 29.0, both below several of the other listed models. Those are Baidu’s numbers. Its launch post puts generation speed above 200 tokens a second. The product pages skip that figure and say only that decoding stays fast.
The Qianfan model is listed as ernie-5.5-flash-preview. Baidu bills in yuan, per thousand tokens. Per million tokens, the promotional rate on Qianfan is 1 yuan for input and 4 yuan for output in peak hours, and 0.5 yuan and 2 yuan off-peak. That’s half the standard list price. If a request hits the cache, Qianfan charges a tenth of that window’s input rate. The billing page doesn’t say when the promotion ends.
| Window | Input, promotional | Input, standard | Output, promotional | Output, standard |
|---|---|---|---|---|
| Peak | 1 | 2 | 4 | 8 |
| Off-peak | 0.5 | 1 | 2 | 4 |
OpenRouter added the model on 10 November, a day before Baidu’s product post, at US$0.28 per million input tokens and US$1.13 per million output tokens. It lists Baidu Qianfan as the provider. Those are OpenRouter’s dollar prices.
The weights still aren’t public. Baidu’s ERNIE 5.5 Flash collection on Hugging Face still lists only the Base checkpoints from 5 November, released under Baidu’s own model licence, which includes territory restrictions. Baidu’s vLLM-ERNIE plugin already has serving notes for the Preview model, including Hugging Face repository names. Those repositories haven’t been published. Baidu has promised Apache 2.0 for the weights once they’re out.
Discover more from Sulat AI & Tech News
Subscribe to get the latest posts sent to your email.
