Baidu Opens ERNIE 5.5 Flash Preview API, Promises Weights by Nov 25

Baidu put ERNIE 5.5 Flash Preview on the market on November 12 at its Baidu World 2026 conference, charging RMB 1 per million input tokens on its Qianfan cloud platform — and promised to open-source the model weights by November 25. The sparse mixture-of-experts model carries 280 billion total parameters with 14 billion activated per token, and it supports a 1-million-token context window. Baidu says it performs strongly in coding, 3D modeling, knowledge work and content creation.

The preview is text-only. According to a third-party aggregator, it doesn't see the screen natively and relies on tools when visual understanding is required. The aggregator also positions the model for long-horizon tasks built on planning, tool use and iterative refinement.

What does this change for developers?

Availability, mostly. The model is live now via API on Qianfan and free on OpenRouter, under the model ID ernie-5.5-flash-preview. Full pricing on Qianfan: RMB 1 per million input tokens, RMB 4 per million output tokens, and RMB 0.04 per million cached input tokens — the cached rate is 25 times cheaper than standard input.

The open-weight promise matters more than the API. A third-party aggregator reports the weights are set for release under Apache 2.0. That's a meaningful shift from the base model's license, and Baidu's own press release and base model blog don't specify the preview model's license. Treat November 25 as a timeline, not a delivery — the weights aren't public yet.

A post-trained sibling to the November 5 base release

ERNIE 5.5 Flash Preview builds on ERNIE 5.5 Flash Base, a pre-training checkpoint Baidu open-sourced on November 5. The base model shares the same 280B/14B architecture. Baidu trained it from scratch on 33.5 trillion tokens with a 128K training context, and it supports 1M tokens for fine-tuning and inference.

A third-party aggregator calls the preview "the first post-trained model built on ERNIE 5.5 Flash Base" and describes it as "an intermediate checkpoint from post-training that is still being improved, and some capabilities may still fall short." That's unusually candid framing — anyone deploying in production should read it as a warning label.

Self-reported benchmarks, with caveats

Baidu says the preview achieves "competitive scores while maintaining extremely fast decoding speed" on Terminal-Bench and AutomationBench.

For the base model, Baidu published a 30-benchmark evaluation against three rivals — DeepSeek-V4-Flash-Base (284B-A13B), Hy3-preview-Base (295B-A21B) and MiMo-V2.5-Base (310B-A15B) — all run through Baidu's unified pipeline with the same prompts and few-shot settings. The base model leads on 24 of 30 benchmarks, including:

• MMLU-Pro: 78.4 vs 69.2
• HumanEval+: 82.3 vs 69.5
• LongBench-v2: 48.7 vs 44.5

It trails on six: ChineseSimpleQA, INCLUDE-lite, HELMET-Rerank, InfiniteBench, BBH (CoT) and FullStackBench. Baidu itself notes that competitor numbers in its pipeline differ slightly from vendors' own reports. Independent evaluations of the preview model don't exist yet.

Inside the architecture

The base model's technical details, already published, explain where the speed comes from. A hybrid attention scheme gives 37 of 43 layers Heavily Compressed Attention (HCA), caching one entry per 128 tokens with 512-dimension heads. The remaining six layers use Multi-head Latent Attention with sparse attention (MLA-DSA) at 256-dimension heads to keep token-level access.

Each MoE layer has 512 routed experts (top-10) plus one shared expert, governed by multiple independent sigmoid routers. Baidu says this "Multi-Router" design improves correlation between routing scores and actual expert utility, especially in middle and deep layers. Experts work in a 2048-dimension latent space — half the hidden width — which cuts dispatch/combine traffic and expert FLOPs roughly in half.

The model also replaces multi-token prediction with ECHO, which reuses the backbone's final layer to predict future tokens and doubles as a built-in draft model for speculative decoding. Pre-training ran on the Muon optimizer with FP8 mixed precision, using a distributed Muon implementation that fuses dense and MoE parameter synchronization across thousands of GPUs.

Baidu says the preview is optimized with ERNIE Evolve, its proprietary recursive self-improvement technology, "to boost inference efficiency and reduce cost." The release doesn't quantify that claim.

Demos target builders

Official demos listed by the aggregator include:

• Game development in Unreal Engine, Unity and Godot
• Procedural 3D work in Blender and FreeCAD
• Research reports and spreadsheets
• Tool-assisted video production

Real-world performance on these tasks hasn't been independently tested.

A licensing split that raises questions

The base model's weights are on Hugging Face and ModelScope in FP8 and BF16 safetensors — but under a custom ERNIE 5.5 Flash Base Model License Agreement v1.0. That license excludes the EU, UK and South Korea, bans military use, forbids using outputs to improve other AI models, and requires commercial users with over US$50 million in related annual revenue to seek authorization.

An Apache 2.0 release for the preview would be far more permissive. That gap between the two licenses, if it holds, would let smaller shops self-host the post-trained model without the base model's restrictions — but only once the weights actually ship.

Part of a broader push

The launch accompanied Baidu's Token Exchange marketplace for model services, upgrades to agents DuMate, KOOKO, Baidu Yijing and Famou 3.0, and hardware announcements including the Tianchi 1024-card Supernode and the Kunlunxin M100/M300 chip roadmap. CEO Robin Li framed the moves around three principles: breaking job boundaries, capturing tacit knowledge as a shared asset, and designing closed-loop processes for continuous evolution.

The preview announcement names no spokesperson and offers no direct leadership quote. If Baidu hits the November 25 date with a true Apache 2.0 release, it plants a competitive open-weight alternative in the 300B-class MoE tier — and the base model's license carve-outs will look like a strategic split, not an oversight.