With ERNIE 5.5 Flash Preview, Baidu released a new language model on 11 November 2026 that is built primarily for efficiency. The model has 280 billion parameters, of which only 14 billion are active per token, and handles up to one million tokens of context. The API is available through Baidu’s Qianfan platform, and the weights are due to follow under the Apache 2.0 licence. Baidu’s own benchmarks paint a mixed picture: ERNIE 5.5 Flash leads in some agent and office tests but trails the top models clearly in terminal coding.
What ERNIE 5.5 Flash Preview is
According to Baidu’s release page, the Preview is the first post-trained model built on ERNIE 5.5 Flash Base, which Baidu published on 5 November. It is a sparse mixture-of-experts (MoE) model: instead of using every parameter for each token, the network activates only a small subset of “experts”. That cuts the compute needed to generate text without shrinking the model as a whole.
The key figures according to Baidu:
- 280 billion parameters in total, 14 billion active per token
- Context window of one million tokens
- Text-only model: it cannot see images on its own
- Focus areas: coding, 3D modelling, knowledge work and content creation
- Status: an intermediate checkpoint from ongoing post-training, not the final version
According to its model card, the base model on Hugging Face was pre-trained on 33.5 trillion tokens with a training context of up to 128,000 tokens. It is available in FP8 and BF16 versions, but under Baidu’s own licence (the “ERNIE 5.5 Flash Base Model License Agreement v1.0”). For the Preview, by contrast, Baidu is promising Apache 2.0. That difference is worth keeping in mind before any commercial use.
Baidu itself stresses that the Preview is not finished. It says the model is already competitive with models of its size, but some capabilities may still fall short. Post-training continues, and improvements are meant to land in the official release.
Pricing and availability
In its press release for the Baidu World 2026 conference, the company gives concrete API prices on Qianfan:
- 1 yuan (RMB) per million input tokens
- 4 yuan per million output tokens
- 0.04 yuan per million tokens of cached input
Baidu adds that the model is also available for free on OpenRouter. The press release says the weights will be open-sourced by 25 November, while the release page only says “soon”. Whether that date holds can only be checked then. Baidu also credits the efficiency to an in-house technique called “ERNIE Evolve”, which it describes as recursive self-improvement. The release contains no solid technical detail on it.
ERNIE 5.5 Flash in the benchmarks
The benchmark table comes from Baidu, and many of the scores for competing models are, by Baidu’s own note, self-evaluated. There are no independent measurements yet. With that caveat, it is notable that Baidu also shows tests in which its own model clearly loses:
- Terminal-Bench 4.0: 46.5 points, well behind Claude Opus 5.5 (63.6) and GPT-6.1 Sol (56.1), but ahead of DeepSeek V4.1 Flash 1017 (40.9)
- DeepSWE v1.1: 71.2, behind DeepSeek V4.1 Flash 1017 (77.0), GPT-6.1 Sol (75.2) and Claude Opus 5.5 (74.2)
- AutomationBench: 61.9, first place ahead of Claude Opus 5.5 (58.0)
- SpreadsheetBench V2: 36.4, narrowly ahead of Claude Opus 5.5 (35.1)
- APEX-Agents 1.1: 64.2, behind Hy4 (75.0) and Claude Opus 5.5 (72.9)
- Agents’ Last Exam (CLI): 29.0, behind DeepSeek V4.1 Flash 1017 (35.2)
That makes the positioning more honest than some launches: ERNIE 5.5 Flash is not a new frontier model but an inexpensive workhorse that keeps up in office and automation tasks. In agentic coding, models such as DeepSeek V4.1-Flash-1017, which competes in the same weight class, sometimes do better even in Baidu’s own table.
What Baidu shows as real-world examples
Most of the release page is taken up by demos that, according to Baidu, the model produced end to end inside an agent harness. Some examples:
- Game development: In a single 11-hour coding task, the model reportedly merged assets from several Unreal Engine generations into one Unreal Engine 5.8 project and wrote the entire game in C++. Because it cannot see the screen, it turned problems found in playtesting into measurable tests in code. Baidu also says it built nearly 100 mobile mini-games from design documents.
- 3D and CAD: Using the Python APIs of Blender and FreeCAD, the model built, among other things, an airport with an airliner, a temple complex and an editable engine model.
- Knowledge work: For domains such as finance and law, Baidu says it built 300,000 training workspaces holding 10 million reference files in more than 40 file types. The demos show research reports and spreadsheet tools.
- Video: More than 150,000 environments for explainers and similar formats are said to have produced about 5,300 minutes of finished video so far.
Demos like these are hand-picked successes. They show what is possible with a suitable agent harness and external tools, not how often it works on the first try. For developers, the text-only point matters: anyone automating visual tasks needs additional tools, such as a separate vision model, which Baidu itself wired in through an MCP tool in one CAD demo.
The setting: Baidu World 2026
The model was presented around Baidu World 2026, where CEO Robin Li focused mainly on AI inside organisations. Alongside ERNIE 5.5 Flash Preview, the press release announces two specialised application models: ERNIE Wizshot Micro-Expression for expressive video and ERNIE Light Deep Thinking for complex requests with tool calls. There is also “Token Exchange”, a marketplace where customers are meant to compare model services from different providers with pricing and performance data. Baidu says more than 100 model resources were listed there after one month. For developers outside China, these announcements are less tangible than the model itself for now, but they show that Baidu sees ERNIE as one building block of a larger platform.
Context
ERNIE 5.5 Flash joins a wave of efficient MoE models from China that court developers with few active parameters and low prices. Similar positioning came recently from Meituan’s LongCat-2.5, and Meta’s small model Muse Ember targets routine agent work in a similar way. ERNIE 5.5 Flash becomes genuinely interesting once the weights actually ship under Apache 2.0. The model could then be self-hosted and adapted without licensing hurdles, keeping in mind that 280 billion parameters need plenty of memory despite sparse activation.
Anyone who wants to try the model can do so through Qianfan or, according to Baidu, for free via OpenRouter. For production decisions, it is worth waiting for the final version and for independent benchmarks.
